nimbus-eth2/beacon_chain/consensus_object_pools/blockchain_dag.nim

2276 lines
84 KiB
Nim
Raw Normal View History

# beacon_chain
# Copyright (c) 2018-2022 Status Research & Development GmbH
# Licensed and distributed under either of
# * MIT license (license terms in the root directory or at https://opensource.org/licenses/MIT).
# * Apache v2 license (license terms in the root directory or at https://www.apache.org/licenses/LICENSE-2.0).
# at your option. This file may not be copied, modified, or distributed except according to those terms.
2022-07-29 10:53:42 +00:00
when (NimMajor, NimMinor) < (1, 4):
{.push raises: [Defect].}
else:
{.push raises: [].}
import
std/[algorithm, options, sequtils, tables, sets],
stew/[assign2, byteutils, results],
metrics, snappy, chronicles,
../spec/[beaconstate, eth2_merkleization, eth2_ssz_serialization, helpers,
disentangle eth2 types from the ssz library (#2785) * reorganize ssz dependencies This PR continues the work in https://github.com/status-im/nimbus-eth2/pull/2646, https://github.com/status-im/nimbus-eth2/pull/2779 as well as past issues with serialization and type, to disentangle SSZ from eth2 and at the same time simplify imports and exports with a structured approach. The principal idea here is that when a library wants to introduce SSZ support, they do so via 3 files: * `ssz_codecs` which imports and reexports `codecs` - this covers the basic byte conversions and ensures no overloads get lost * `xxx_merkleization` imports and exports `merkleization` to specialize and get access to `hash_tree_root` and friends * `xxx_ssz_serialization` imports and exports `ssz_serialization` to specialize ssz for a specific library Those that need to interact with SSZ always import the `xxx_` versions of the modules and never `ssz` itself so as to keep imports simple and safe. This is similar to how the REST / JSON-RPC serializers are structured in that someone wanting to serialize spec types to REST-JSON will import `eth2_rest_serialization` and nothing else. * split up ssz into a core library that is independendent of eth2 types * rename `bytes_reader` to `codec` to highlight that it contains coding and decoding of bytes and native ssz types * remove tricky List init overload that causes compile issues * get rid of top-level ssz import * reenable merkleization tests * move some "standard" json serializers to spec * remove `ValidatorIndex` serialization for now * remove test_ssz_merkleization * add tests for over/underlong byte sequences * fix broken seq[byte] test - seq[byte] is not an SSZ type There are a few things this PR doesn't solve: * like #2646 this PR is weak on how to handle root and other dontSerialize fields that "sometimes" should be computed - the same problem appears in REST / JSON-RPC etc * Fix a build problem on macOS * Another way to fix the macOS builds Co-authored-by: Zahary Karadjov <zahary@gmail.com>
2021-08-18 18:57:58 +00:00
state_transition, validator],
../spec/datatypes/[phase0, altair, bellatrix],
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
".."/[beacon_chain_db, era_db],
"."/[block_pools_types, block_quarantine]
from ../spec/datatypes/capella import shortLog
export
eth2_merkleization, eth2_ssz_serialization,
block_pools_types, results, beacon_chain_db
# https://github.com/ethereum/beacon-metrics/blob/master/metrics.md#interop-metrics
declareGauge beacon_head_root, "Root of the head block of the beacon chain"
declareGauge beacon_head_slot, "Slot of the head block of the beacon chain"
# https://github.com/ethereum/beacon-metrics/blob/master/metrics.md#interop-metrics
declareGauge beacon_finalized_epoch, "Current finalized epoch" # On epoch transition
declareGauge beacon_finalized_root, "Current finalized root" # On epoch transition
declareGauge beacon_current_justified_epoch, "Current justified epoch" # On epoch transition
declareGauge beacon_current_justified_root, "Current justified root" # On epoch transition
declareGauge beacon_previous_justified_epoch, "Current previously justified epoch" # On epoch transition
declareGauge beacon_previous_justified_root, "Current previously justified root" # On epoch transition
declareGauge beacon_reorgs_total_total, "Total occurrences of reorganizations of the chain" # On fork choice; backwards-compat name (used to be a counter)
declareGauge beacon_reorgs_total, "Total occurrences of reorganizations of the chain" # Interop copy
declareCounter beacon_state_data_cache_hits, "EpochRef hits"
declareCounter beacon_state_data_cache_misses, "EpochRef misses"
declareCounter beacon_state_rewinds, "State database rewinds"
declareGauge beacon_active_validators, "Number of validators in the active validator set"
declareGauge beacon_current_active_validators, "Number of validators in the active validator set" # Interop copy
declareGauge beacon_pending_deposits, "Number of pending deposits (state.eth1_data.deposit_count - state.eth1_deposit_index)" # On block
declareGauge beacon_processed_deposits_total, "Number of total deposits included on chain" # On block
logScope: topics = "chaindag"
const
# When finality happens, we prune historical states from the database except
# for a snapshort every 32 epochs from which replays can happen - there's a
# balance here between making long replays and saving on disk space
EPOCHS_PER_STATE_SNAPSHOT = 32
proc putBlock*(
dag: ChainDAGRef, signedBlock: ForkyTrustedSignedBeaconBlock) =
dag.db.putBlock(signedBlock)
proc updateState*(
dag: ChainDAGRef, state: var ForkedHashedBeaconState, bsi: BlockSlotId,
save: bool, cache: var StateCache): bool {.gcsafe.}
template withUpdatedState*(
dag: ChainDAGRef, stateParam: var ForkedHashedBeaconState,
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
bsiParam: BlockSlotId, okBody: untyped, failureBody: untyped): untyped =
## Helper template that updates stateData to a particular BlockSlot - usage of
## stateData is unsafe outside of block, or across `await` boundaries
block:
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let bsi {.inject.} = bsiParam
var cache {.inject.} = StateCache()
if updateState(dag, stateParam, bsi, false, cache):
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
template bid(): BlockId {.inject, used.} = bsi.bid
template updatedState(): ForkedHashedBeaconState {.inject, used.} = stateParam
okBody
else:
failureBody
func get_effective_balances(validators: openArray[Validator], epoch: Epoch):
seq[Gwei] =
## Get the balances from a state as counted for fork choice
result.newSeq(validators.len) # zero-init
for i in 0 ..< result.len:
# All non-active validators have a 0 balance
let validator = unsafeAddr validators[i]
performance fixes (#2259) * performance fixes * don't mark tree cache as dirty on read-only List accesses * store only blob in memory for keys and signatures, parse blob lazily * compare public keys by blob instead of parsing / converting to raw * compare Eth2Digest using non-constant-time comparison * avoid some unnecessary validator copying This branch will in particular speed up deposit processing which has been slowing down block replay. Pre (mainnet, 1600 blocks): ``` All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 3450.269, 0.000, 3450.269, 3450.269, 1, Initialize DB 0.417, 0.822, 0.036, 21.098, 1400, Load block from database 16.521, 0.000, 16.521, 16.521, 1, Load state from database 27.906, 50.846, 8.104, 1507.633, 1350, Apply block 52.617, 37.029, 20.640, 135.938, 50, Apply epoch block ``` Post: ``` 3502.715, 0.000, 3502.715, 3502.715, 1, Initialize DB 0.080, 0.560, 0.035, 21.015, 1400, Load block from database 17.595, 0.000, 17.595, 17.595, 1, Load state from database 15.706, 11.028, 8.300, 107.537, 1350, Apply block 33.217, 12.622, 17.331, 60.580, 50, Apply epoch block ``` * more perf fixes * load EpochRef cache into StateCache more aggressively * point out security concern with public key cache * reuse proposer index from state when processing block * avoid genericAssign in a few more places * don't parse key when signature is unparseable * fix `==` overload for Eth2Digest * preallocate validator list when getting active validators * speed up proposer index calculation a little bit * reuse cache when replaying blocks in ncli_db * avoid a few more copying loops ``` Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 3279.158, 0.000, 3279.158, 3279.158, 1, Initialize DB 0.072, 0.357, 0.035, 13.400, 1400, Load block from database 17.295, 0.000, 17.295, 17.295, 1, Load state from database 5.918, 9.896, 0.198, 98.028, 1350, Apply block 15.888, 10.951, 7.902, 39.535, 50, Apply epoch block 0.000, 0.000, 0.000, 0.000, 0, Database block store ``` * clear full balance cache before processing rewards and penalties ``` All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 3947.901, 0.000, 3947.901, 3947.901, 1, Initialize DB 0.124, 0.506, 0.026, 202.370, 363345, Load block from database 97.614, 0.000, 97.614, 97.614, 1, Load state from database 0.186, 0.188, 0.012, 99.561, 357262, Advance slot, non-epoch 14.161, 5.966, 1.099, 395.511, 11524, Advance slot, epoch 1.372, 4.170, 0.017, 276.401, 363345, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database block store ```
2021-01-25 12:04:18 +00:00
if validator[].is_active_validator(epoch):
result[i] = validator[].effective_balance
proc updateValidatorKeys*(dag: ChainDAGRef, validators: openArray[Validator]) =
# Update validator key cache - must be called every time a valid block is
# applied to the state - this is important to ensure that when we sync blocks
# without storing a state (non-epoch blocks essentially), the deposits from
# those blocks are persisted to the in-database cache of immutable validator
# data (but no earlier than that the whole block as been validated)
dag.db.updateImmutableValidators(validators)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
proc updateFinalizedBlocks*(db: BeaconChainDB, newFinalized: openArray[BlockId]) =
if db.db.readOnly: return # TODO abstraction leak - where to put this?
chaindag: don't keep backfill block table in memory (#3429) This PR names and documents the concept of the archive: a range of slots for which we have degraded functionality in terms of historical access - in particular: * we don't support rewinding to states in this range * we don't keep an in-memory representation of the block dag The archive de-facto exists in a trusted-node-synced node, but this PR gives it a name and drops the in-memory digest index. In order to satisfy `GetBlocksByRange` requests, we ensure that we have blocks for the entire archive period via backfill. Future versions may relax this further, adding a "pre-archive" period that is fully pruned. During by-slot searches in the archive (both for libp2p and rest requests), an extra database lookup is used to covert the given `slot` to a `root` - future versions will avoid this using era files which natively are indexed by `slot`. That said, the lookup is quite fast compared to the actual block loading given how trivial the table is - it's hard to measure, even. A collateral benefit of this PR is that checkpoint-synced nodes will see 100-200MB memory usage savings, thanks to the dropped in-memory cache - future pruning work will bring this benefit to full nodes as well. * document chaindag storage architecture and assumptions * look up parent using block id instead of full block in clearance (future-proofing the code against a future in which blocks come from era files) * simplify finalized block init, always writing the backfill portion to db at startup (to ensure lookups work as expected) * preallocate some extra memory for finalized blocks, to avoid immediate realloc
2022-02-26 18:16:19 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
db.withManyWrites:
for bid in newFinalized:
db.finalizedBlocks.insert(bid.slot, bid.root)
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
proc updateFrontfillBlocks*(dag: ChainDAGRef) =
# When backfilling is done and manages to reach the frontfill point, we can
# write the frontfill index knowing that the block information in the
# era files match the chain
if dag.db.db.readOnly: return # TODO abstraction leak - where to put this?
if dag.frontfillBlocks.len == 0 or dag.backfill.slot > 0:
return
info "Writing frontfill index", slots = dag.frontfillBlocks.len
dag.db.withManyWrites:
let low = dag.db.finalizedBlocks.low.expect(
"wrote at least tailRef during init")
let blocks = min(low.int, dag.frontfillBlocks.len - 1)
var parent: Eth2Digest
for i in 0..blocks:
let root = dag.frontfillBlocks[i]
if not isZero(root):
dag.db.finalizedBlocks.insert(Slot(i), root)
dag.db.putBeaconBlockSummary(
root, BeaconBlockSummary(slot: Slot(i), parent_root: parent))
parent = root
reset(dag.frontfillBlocks)
func validatorKey*(
dag: ChainDAGRef, index: ValidatorIndex or uint64): Option[CookedPubKey] =
## Returns the validator pubkey for the index, assuming it's been observed
## at any point in time - this function may return pubkeys for indicies that
## are not (yet) part of the head state (if the key has been observed on a
## non-head branch)!
dag.db.immutableValidators.load(index)
template is_merge_transition_complete*(
stateParam: ForkedHashedBeaconState): bool =
withState(stateParam):
when stateFork >= BeaconStateFork.Bellatrix:
is_merge_transition_complete(forkyState.data)
else:
false
func effective_balances*(epochRef: EpochRef): seq[Gwei] =
try:
SSZ.decode(snappy.decode(epochRef.effective_balances_bytes, uint32.high),
List[Gwei, Limit VALIDATOR_REGISTRY_LIMIT]).toSeq()
except CatchableError as exc:
raiseAssert exc.msg
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
func getBlockRef*(dag: ChainDAGRef, root: Eth2Digest): Opt[BlockRef] =
## Retrieve a resolved block reference, if available - this function does
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
## not return historical finalized blocks, see `getBlockIdAtSlot` for a
## function that covers the entire known history
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
let key = KeyedBlockRef.asLookupKey(root)
# HashSet lacks the api to do check-and-get in one lookup - `[]` will return
# the copy of the instance in the set which has more fields than `root` set!
if key in dag.forkBlocks:
try: ok(dag.forkBlocks[key].blockRef())
except KeyError: raiseAssert "contains"
else:
err()
func getBlockIdAtSlot*(dag: ChainDAGRef, slot: Slot): Opt[BlockSlotId] =
## Retrieve the canonical block at the given slot, or the last block that
chaindag: don't keep backfill block table in memory (#3429) This PR names and documents the concept of the archive: a range of slots for which we have degraded functionality in terms of historical access - in particular: * we don't support rewinding to states in this range * we don't keep an in-memory representation of the block dag The archive de-facto exists in a trusted-node-synced node, but this PR gives it a name and drops the in-memory digest index. In order to satisfy `GetBlocksByRange` requests, we ensure that we have blocks for the entire archive period via backfill. Future versions may relax this further, adding a "pre-archive" period that is fully pruned. During by-slot searches in the archive (both for libp2p and rest requests), an extra database lookup is used to covert the given `slot` to a `root` - future versions will avoid this using era files which natively are indexed by `slot`. That said, the lookup is quite fast compared to the actual block loading given how trivial the table is - it's hard to measure, even. A collateral benefit of this PR is that checkpoint-synced nodes will see 100-200MB memory usage savings, thanks to the dropped in-memory cache - future pruning work will bring this benefit to full nodes as well. * document chaindag storage architecture and assumptions * look up parent using block id instead of full block in clearance (future-proofing the code against a future in which blocks come from era files) * simplify finalized block init, always writing the backfill portion to db at startup (to ensure lookups work as expected) * preallocate some extra memory for finalized blocks, to avoid immediate realloc
2022-02-26 18:16:19 +00:00
## comes before - similar to atSlot, but without the linear scan - may hit
## the database to look up early indices.
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if slot > dag.finalizedHead.slot:
return dag.head.atSlot(slot).toBlockSlotId() # iterate to the given slot
if dag.finalizedHead.blck == nil:
# Not initialized yet (in init)
return Opt.none(BlockSlotId)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if slot >= dag.finalizedHead.blck.slot:
# finalized head is still in memory
return dag.finalizedHead.blck.atSlot(slot).toBlockSlotId()
chaindag: don't keep backfill block table in memory (#3429) This PR names and documents the concept of the archive: a range of slots for which we have degraded functionality in terms of historical access - in particular: * we don't support rewinding to states in this range * we don't keep an in-memory representation of the block dag The archive de-facto exists in a trusted-node-synced node, but this PR gives it a name and drops the in-memory digest index. In order to satisfy `GetBlocksByRange` requests, we ensure that we have blocks for the entire archive period via backfill. Future versions may relax this further, adding a "pre-archive" period that is fully pruned. During by-slot searches in the archive (both for libp2p and rest requests), an extra database lookup is used to covert the given `slot` to a `root` - future versions will avoid this using era files which natively are indexed by `slot`. That said, the lookup is quite fast compared to the actual block loading given how trivial the table is - it's hard to measure, even. A collateral benefit of this PR is that checkpoint-synced nodes will see 100-200MB memory usage savings, thanks to the dropped in-memory cache - future pruning work will bring this benefit to full nodes as well. * document chaindag storage architecture and assumptions * look up parent using block id instead of full block in clearance (future-proofing the code against a future in which blocks come from era files) * simplify finalized block init, always writing the backfill portion to db at startup (to ensure lookups work as expected) * preallocate some extra memory for finalized blocks, to avoid immediate realloc
2022-02-26 18:16:19 +00:00
let finlow = dag.db.finalizedBlocks.low.expect("at least tailRef written")
if slot >= finlow:
var pos = slot
while true:
let root = dag.db.finalizedBlocks.get(pos)
if root.isSome():
return ok BlockSlotId.init(
BlockId(root: root.get(), slot: pos), slot)
chaindag: don't keep backfill block table in memory (#3429) This PR names and documents the concept of the archive: a range of slots for which we have degraded functionality in terms of historical access - in particular: * we don't support rewinding to states in this range * we don't keep an in-memory representation of the block dag The archive de-facto exists in a trusted-node-synced node, but this PR gives it a name and drops the in-memory digest index. In order to satisfy `GetBlocksByRange` requests, we ensure that we have blocks for the entire archive period via backfill. Future versions may relax this further, adding a "pre-archive" period that is fully pruned. During by-slot searches in the archive (both for libp2p and rest requests), an extra database lookup is used to covert the given `slot` to a `root` - future versions will avoid this using era files which natively are indexed by `slot`. That said, the lookup is quite fast compared to the actual block loading given how trivial the table is - it's hard to measure, even. A collateral benefit of this PR is that checkpoint-synced nodes will see 100-200MB memory usage savings, thanks to the dropped in-memory cache - future pruning work will bring this benefit to full nodes as well. * document chaindag storage architecture and assumptions * look up parent using block id instead of full block in clearance (future-proofing the code against a future in which blocks come from era files) * simplify finalized block init, always writing the backfill portion to db at startup (to ensure lookups work as expected) * preallocate some extra memory for finalized blocks, to avoid immediate realloc
2022-02-26 18:16:19 +00:00
doAssert pos > finlow, "We should have returned the finlow"
pos = pos - 1
if slot == GENESIS_SLOT and dag.genesis.isSome():
return ok dag.genesis.get().atSlot()
err() # not backfilled yet
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
proc containsBlock(
cfg: RuntimeConfig, db: BeaconChainDB, slot: Slot, root: Eth2Digest): bool =
db.containsBlock(root, cfg.blockForkAtEpoch(slot.epoch))
proc getForkedBlock*(db: BeaconChainDB, root: Eth2Digest):
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
Opt[ForkedTrustedSignedBeaconBlock] =
# When we only have a digest, we don't know which fork it's from so we try
# them one by one - this should be used sparingly
if (let blck = db.getBlock(root, bellatrix.TrustedSignedBeaconBlock);
blck.isSome()):
ok(ForkedTrustedSignedBeaconBlock.init(blck.get()))
elif (let blck = db.getBlock(root, altair.TrustedSignedBeaconBlock);
blck.isSome()):
ok(ForkedTrustedSignedBeaconBlock.init(blck.get()))
elif (let blck = db.getBlock(root, phase0.TrustedSignedBeaconBlock);
blck.isSome()):
ok(ForkedTrustedSignedBeaconBlock.init(blck.get()))
else:
err()
proc containsBlock(dag: ChainDAGRef, bid: BlockId): bool =
let fork = dag.cfg.blockForkAtEpoch(bid.slot.epoch)
if dag.db.containsBlock(bid.root, fork):
return true
# TODO avoid loading bytes from era
var bytes: seq[byte]
(bid.slot <= dag.finalizedHead.slot and
getBlockSZ(
dag.era, getStateField(dag.headState, historical_roots).asSeq,
bid.slot, bytes).isOk and bytes.len > 0)
proc getBlock*(
dag: ChainDAGRef, bid: BlockId,
T: type ForkyTrustedSignedBeaconBlock): Opt[T] =
dag.db.getBlock(bid.root, T) or
getBlock(
dag.era, getStateField(dag.headState, historical_roots).asSeq,
bid.slot, Opt[Eth2Digest].ok(bid.root), T)
proc getBlockSSZ*(dag: ChainDAGRef, bid: BlockId, bytes: var seq[byte]): bool =
# Load the SSZ-encoded data of a block into `bytes`, overwriting the existing
# content
let fork = dag.cfg.blockForkAtEpoch(bid.slot.epoch)
dag.db.getBlockSSZ(bid.root, bytes, fork) or
(bid.slot <= dag.finalizedHead.slot and
getBlockSSZ(
dag.era, getStateField(dag.headState, historical_roots).asSeq,
bid.slot, bytes).isOk)
proc getBlockSZ*(dag: ChainDAGRef, bid: BlockId, bytes: var seq[byte]): bool =
# Load the snappy-frame-compressed ("SZ") SSZ-encoded data of a block into
# `bytes`, overwriting the existing content
# careful: there are two snappy encodings in use, with and without framing!
# Returns true if the block is found, false if not
let fork = dag.cfg.blockForkAtEpoch(bid.slot.epoch)
dag.db.getBlockSZ(bid.root, bytes, fork) or
(bid.slot <= dag.finalizedHead.slot and
getBlockSZ(
dag.era, getStateField(dag.headState, historical_roots).asSeq,
bid.slot, bytes).isOk)
proc getForkedBlock*(
dag: ChainDAGRef, bid: BlockId): Opt[ForkedTrustedSignedBeaconBlock] =
let fork = dag.cfg.blockForkAtEpoch(bid.slot.epoch)
result.ok(ForkedTrustedSignedBeaconBlock(kind: fork))
withBlck(result.get()):
type T = type(blck)
blck = getBlock(dag, bid, T).valueOr:
getBlock(
dag.era, getStateField(dag.headState, historical_roots).asSeq,
bid.slot, Opt[Eth2Digest].ok(bid.root), T).valueOr:
result.err()
return
proc getBlockId*(db: BeaconChainDB, root: Eth2Digest): Opt[BlockId] =
block: # We might have a summary in the database
let summary = db.getBeaconBlockSummary(root)
if summary.isOk():
return ok(BlockId(root: root, slot: summary.get().slot))
block:
# We might have a block without having written a summary - this can happen
# if there was a crash between writing the block and writing the summary,
# specially in databases written by older nimbus versions
let forked = db.getForkedBlock(root)
if forked.isSome():
# Shouldn't happen too often but..
let
blck = forked.get()
summary = withBlck(blck): blck.message.toBeaconBlockSummary()
debug "Writing summary", blck = shortLog(blck)
db.putBeaconBlockSummary(root, summary)
return ok(BlockId(root: root, slot: summary.slot))
err()
proc getBlockId*(dag: ChainDAGRef, root: Eth2Digest): Opt[BlockId] =
chaindag: don't keep backfill block table in memory (#3429) This PR names and documents the concept of the archive: a range of slots for which we have degraded functionality in terms of historical access - in particular: * we don't support rewinding to states in this range * we don't keep an in-memory representation of the block dag The archive de-facto exists in a trusted-node-synced node, but this PR gives it a name and drops the in-memory digest index. In order to satisfy `GetBlocksByRange` requests, we ensure that we have blocks for the entire archive period via backfill. Future versions may relax this further, adding a "pre-archive" period that is fully pruned. During by-slot searches in the archive (both for libp2p and rest requests), an extra database lookup is used to covert the given `slot` to a `root` - future versions will avoid this using era files which natively are indexed by `slot`. That said, the lookup is quite fast compared to the actual block loading given how trivial the table is - it's hard to measure, even. A collateral benefit of this PR is that checkpoint-synced nodes will see 100-200MB memory usage savings, thanks to the dropped in-memory cache - future pruning work will bring this benefit to full nodes as well. * document chaindag storage architecture and assumptions * look up parent using block id instead of full block in clearance (future-proofing the code against a future in which blocks come from era files) * simplify finalized block init, always writing the backfill portion to db at startup (to ensure lookups work as expected) * preallocate some extra memory for finalized blocks, to avoid immediate realloc
2022-02-26 18:16:19 +00:00
## Look up block id by root in history - useful for turning a root into a
## slot - may hit the database, may return blocks that have since become
## unviable - use `getBlockIdAtSlot` to check that the block is still viable
## if used in a sensitive context
block: # If we have a BlockRef, this is the fastest way to get a block id
let blck = dag.getBlockRef(root)
if blck.isOk():
return ok(blck.get().bid)
dag.db.getBlockId(root)
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
proc getForkedBlock*(
dag: ChainDAGRef, root: Eth2Digest): Opt[ForkedTrustedSignedBeaconBlock] =
let bid = dag.getBlockId(root)
if bid.isSome():
dag.getForkedBlock(bid.get())
else:
# In case we didn't have a summary - should be rare, but ..
dag.db.getForkedBlock(root)
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
func isCanonical*(dag: ChainDAGRef, bid: BlockId): bool =
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
## Return true iff the given `bid` is part of the history selected by `dag.head`
let current = dag.getBlockIdAtSlot(bid.slot).valueOr:
return false # We don't know, so ..
return current.bid == bid
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
func parent*(dag: ChainDAGRef, bid: BlockId): Opt[BlockId] =
if bid.slot == 0:
return err()
if bid.slot > dag.finalizedHead.slot:
# Make sure we follow the correct history as there may be forks
let blck = ? dag.getBlockRef(bid.root)
doAssert not isNil(blck.parent), "should reach finalized head"
return ok blck.parent.bid
let bids = ? dag.getBlockIdAtSlot(bid.slot - 1)
ok(bids.bid)
func parentOrSlot*(dag: ChainDAGRef, bsi: BlockSlotId): Opt[BlockSlotId] =
if bsi.slot == 0:
return err()
if bsi.isProposed:
let parent = ? dag.parent(bsi.bid)
ok BlockSlotId.init(parent, bsi.slot)
else:
ok BlockSlotId.init(bsi.bid, bsi.slot - 1)
func atSlot*(dag: ChainDAGRef, bid: BlockId, slot: Slot): Opt[BlockSlotId] =
if bid.slot > dag.finalizedHead.slot:
let blck = ? dag.getBlockRef(bid.root)
if slot > dag.finalizedHead.slot:
return blck.atSlot(slot).toBlockSlotId()
else:
# Check if the given `bid` is still part of history - it might hail from an
# orphaned fork
let existing = ? dag.getBlockIdAtSlot(bid.slot)
if existing.bid != bid:
return err() # Not part of known / relevant history
if existing.slot == slot: # and bid.slot == slot
return ok existing
if bid.slot <= slot:
ok BlockSlotId.init(bid, slot)
else:
dag.getBlockIdAtSlot(slot)
func nextTimestamp[I, T](cache: var LRUCache[I, T]): uint32 =
if cache.timestamp == uint32.high:
for i in 0 ..< I:
template e: untyped = cache.entries[i]
if e.lastUsed != 0:
e.lastUsed = 1
cache.timestamp = 1
inc cache.timestamp
cache.timestamp
template findIt[I, T](cache: var LRUCache[I, T], predicate: untyped): Opt[T] =
block:
var res: Opt[T]
for i in 0 ..< I:
template e: untyped = cache.entries[i]
template it: untyped {.inject, used.} = e.value
if e.lastUsed != 0 and predicate:
e.lastUsed = cache.nextTimestamp
res.ok it
break
res
template delIt[I, T](cache: var LRUCache[I, T], predicate: untyped) =
block:
for i in 0 ..< I:
template e: untyped = cache.entries[i]
template it: untyped {.inject, used.} = e.value
if e.lastUsed != 0 and predicate:
e.reset()
func put[I, T](cache: var LRUCache[I, T], value: T) =
var lru = 0
block:
var min = uint32.high
for i in 0 ..< I:
template e: untyped = cache.entries[i]
if e.lastUsed < min:
min = e.lastUsed
lru = i
if min == 0:
break
template e: untyped = cache.entries[lru]
e.value = value
e.lastUsed = cache.nextTimestamp
func epochAncestor(dag: ChainDAGRef, bid: BlockId, epoch: Epoch):
Opt[BlockSlotId] =
## The epoch ancestor is the last block that has an effect on the epoch-
## related state data, as updated in `process_epoch` - this block determines
## effective balances, validator addtions and removals etc and serves as a
## base for `EpochRef` construction.
if epoch < dag.tail.slot.epoch or bid.slot < dag.tail.slot:
# Not enough information in database to meaningfully process pre-tail epochs
return Opt.none BlockSlotId
let
dependentSlot =
if epoch == dag.tail.slot.epoch:
# Use the tail as "dependent block" - this may be the genesis block, or,
# in the case of checkpoint sync, the checkpoint block
dag.tail.slot
else:
epoch.start_slot() - 1
bsi = ? dag.atSlot(bid, dependentSlot)
epochSlot =
if epoch == dag.tail.slot.epoch:
dag.tail.slot
else:
epoch.start_slot()
ok BlockSlotId(bid: bsi.bid, slot: epochSlot)
func epochKey(dag: ChainDAGRef, bid: BlockId, epoch: Epoch): Opt[EpochKey] =
## The state transition works by storing information from blocks in a
## "working" area until the epoch transition, then batching work collected
## during the epoch. Thus, last block in the ancestor epochs is the block
## that has an impact on epoch currently considered.
##
## This function returns an epoch key pointing to that epoch boundary, i.e. the
## boundary where the last block has been applied to the state and epoch
## processing has been done.
let bsi = dag.epochAncestor(bid, epoch).valueOr:
return Opt.none(EpochKey)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
Opt.some(EpochKey(bid: bsi.bid, epoch: epoch))
func findShufflingRef*(
dag: ChainDAGRef, bid: BlockId, epoch: Epoch): Opt[ShufflingRef] =
## Lookup a shuffling in the cache, returning `none` if it's not present - see
## `getShufflingRef` for a version that creates a new instance if it's missing
let
dependent_slot = if epoch > 2: (epoch - 1).start_slot() - 1 else: Slot(0)
dependent_bsi = dag.atSlot(bid, dependent_slot).valueOr:
return Opt.none(ShufflingRef)
dag.shufflingRefs.findIt(
it.epoch == epoch and dependent_bsi.bid.root == it.attester_dependent_root)
func putShufflingRef*(dag: ChainDAGRef, shufflingRef: ShufflingRef) =
## Store shuffling in the cache
if shufflingRef.epoch < dag.finalizedHead.slot.epoch():
# Only cache epoch information for unfinalized blocks - earlier states
# are seldomly used (ie RPC), so no need to cache
return
dag.shufflingRefs.put shufflingRef
func findEpochRef*(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag: ChainDAGRef, bid: BlockId, epoch: Epoch): Opt[EpochRef] =
## Lookup an EpochRef in the cache, returning `none` if it's not present - see
## `getEpochRef` for a version that creates a new instance if it's missing
let key = ? dag.epochKey(bid, epoch)
dag.epochRefs.findIt(it.key == key)
func putEpochRef(dag: ChainDAGRef, epochRef: EpochRef) =
if epochRef.epoch < dag.finalizedHead.slot.epoch():
# Only cache epoch information for unfinalized blocks - earlier states
# are seldomly used (ie RPC), so no need to cache
return
dag.epochRefs.put epochRef
func init*(
T: type ShufflingRef, state: ForkedHashedBeaconState,
cache: var StateCache, epoch: Epoch): T =
let
dependent_epoch =
if epoch < 1: Epoch(0) else: epoch - 1
attester_dependent_root =
withState(state): forkyState.dependent_root(dependent_epoch)
ShufflingRef(
epoch: epoch,
attester_dependent_root: attester_dependent_root,
shuffled_active_validator_indices:
cache.get_shuffled_active_validator_indices(state, epoch),
)
func init*(
T: type EpochRef, dag: ChainDAGRef, state: ForkedHashedBeaconState,
cache: var StateCache): T =
let
epoch = state.get_current_epoch()
proposer_dependent_root = withState(state):
forkyState.proposer_dependent_root
shufflingRef = dag.findShufflingRef(state.latest_block_id, epoch).valueOr:
let tmp = ShufflingRef.init(state, cache, epoch)
dag.putShufflingRef(tmp)
tmp
attester_dependent_root = withState(state):
forkyState.attester_dependent_root
epochRef = EpochRef(
key: dag.epochKey(state.latest_block_id, epoch).expect(
"Valid epoch ancestor when processing state"),
eth1_data:
getStateField(state, eth1_data),
eth1_deposit_index:
getStateField(state, eth1_deposit_index),
checkpoints:
FinalityCheckpoints(
justified: getStateField(state, current_justified_checkpoint),
finalized: getStateField(state, finalized_checkpoint)),
# beacon_proposers: Separately filled below
proposer_dependent_root: proposer_dependent_root,
shufflingRef: shufflingRef
)
epochStart = epoch.start_slot()
for i in 0'u64..<SLOTS_PER_EPOCH:
epochRef.beacon_proposers[i] =
get_beacon_proposer_index(state, cache, epochStart + i)
# When fork choice runs, it will need the effective balance of the justified
# checkpoint - we pre-load the balances here to avoid rewinding the justified
# state later and compress them because not all checkpoints end up being used
# for fork choice - specially during long periods of non-finalization
proc snappyEncode(inp: openArray[byte]): seq[byte] =
try:
snappy.encode(inp)
except CatchableError as err:
raiseAssert err.msg
epochRef.effective_balances_bytes =
snappyEncode(SSZ.encode(
List[Gwei, Limit VALIDATOR_REGISTRY_LIMIT](
get_effective_balances(getStateField(state, validators).asSeq, epoch))))
epochRef
performance fixes (#2259) * performance fixes * don't mark tree cache as dirty on read-only List accesses * store only blob in memory for keys and signatures, parse blob lazily * compare public keys by blob instead of parsing / converting to raw * compare Eth2Digest using non-constant-time comparison * avoid some unnecessary validator copying This branch will in particular speed up deposit processing which has been slowing down block replay. Pre (mainnet, 1600 blocks): ``` All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 3450.269, 0.000, 3450.269, 3450.269, 1, Initialize DB 0.417, 0.822, 0.036, 21.098, 1400, Load block from database 16.521, 0.000, 16.521, 16.521, 1, Load state from database 27.906, 50.846, 8.104, 1507.633, 1350, Apply block 52.617, 37.029, 20.640, 135.938, 50, Apply epoch block ``` Post: ``` 3502.715, 0.000, 3502.715, 3502.715, 1, Initialize DB 0.080, 0.560, 0.035, 21.015, 1400, Load block from database 17.595, 0.000, 17.595, 17.595, 1, Load state from database 15.706, 11.028, 8.300, 107.537, 1350, Apply block 33.217, 12.622, 17.331, 60.580, 50, Apply epoch block ``` * more perf fixes * load EpochRef cache into StateCache more aggressively * point out security concern with public key cache * reuse proposer index from state when processing block * avoid genericAssign in a few more places * don't parse key when signature is unparseable * fix `==` overload for Eth2Digest * preallocate validator list when getting active validators * speed up proposer index calculation a little bit * reuse cache when replaying blocks in ncli_db * avoid a few more copying loops ``` Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 3279.158, 0.000, 3279.158, 3279.158, 1, Initialize DB 0.072, 0.357, 0.035, 13.400, 1400, Load block from database 17.295, 0.000, 17.295, 17.295, 1, Load state from database 5.918, 9.896, 0.198, 98.028, 1350, Apply block 15.888, 10.951, 7.902, 39.535, 50, Apply epoch block 0.000, 0.000, 0.000, 0.000, 0, Database block store ``` * clear full balance cache before processing rewards and penalties ``` All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 3947.901, 0.000, 3947.901, 3947.901, 1, Initialize DB 0.124, 0.506, 0.026, 202.370, 363345, Load block from database 97.614, 0.000, 97.614, 97.614, 1, Load state from database 0.186, 0.188, 0.012, 99.561, 357262, Advance slot, non-epoch 14.161, 5.966, 1.099, 395.511, 11524, Advance slot, epoch 1.372, 4.170, 0.017, 276.401, 363345, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database block store ```
2021-01-25 12:04:18 +00:00
func loadStateCache(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag: ChainDAGRef, cache: var StateCache, bid: BlockId, epoch: Epoch) =
# When creating a state cache, we want the current and the previous epoch
# information to be preloaded as both of these are used in state transition
# functions
template load(e: Epoch) =
block:
let epoch = e
if epoch notin cache.shuffled_active_validator_indices:
let shufflingRef = dag.findShufflingRef(bid, epoch)
if shufflingRef.isSome():
cache.shuffled_active_validator_indices[epoch] =
shufflingRef[][].shuffled_active_validator_indices
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let epochRef = dag.findEpochRef(bid, epoch)
if epochRef.isSome():
let start_slot = epoch.start_slot()
for i, idx in epochRef[][].beacon_proposers:
cache.beacon_proposer_indices[start_slot + i] = idx
load(epoch)
if epoch > 0:
load(epoch - 1)
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
func containsForkBlock*(dag: ChainDAGRef, root: Eth2Digest): bool =
## Checks for blocks at the finalized checkpoint or newer
KeyedBlockRef.asLookupKey(root) in dag.forkBlocks
func isFinalizedStateSnapshot(slot: Slot): bool =
slot.is_epoch and slot.epoch mod EPOCHS_PER_STATE_SNAPSHOT == 0
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
func isStateCheckpoint(dag: ChainDAGRef, bsi: BlockSlotId): bool =
## State checkpoints are the points in time for which we store full state
## snapshots, which later serve as rewind starting points when replaying state
## transitions from database, for example during reorgs.
##
# As a policy, we only store epoch boundary states without the epoch block
# (if it exists) applied - the rest can be reconstructed by loading an epoch
# boundary state and applying the missing blocks.
# We also avoid states that were produced with empty slots only - as such,
# there is only a checkpoint for the first epoch after a block.
# The tail block also counts as a state checkpoint!
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
(bsi.isProposed and bsi.bid == dag.tail) or
(bsi.slot.is_epoch and bsi.slot.epoch == (bsi.bid.slot.epoch + 1))
proc getState(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
db: BeaconChainDB, cfg: RuntimeConfig, block_root: Eth2Digest, slot: Slot,
state: var ForkedHashedBeaconState, rollback: RollbackProc): bool =
let state_root = db.getStateRoot(block_root, slot).valueOr:
return false
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
db.getState(cfg.stateForkAtEpoch(slot.epoch), state_root, state, rollback)
proc getState*(
db: BeaconChainDB, cfg: RuntimeConfig, block_root: Eth2Digest,
slots: Slice[Slot], state: var ForkedHashedBeaconState,
rollback: RollbackProc): bool =
var slot = slots.b
while slot >= slots.a:
let state_root = db.getStateRoot(block_root, slot)
if state_root.isSome() and
db.getState(
cfg.stateForkAtEpoch(slot.epoch), state_root.get(), state, rollback):
return true
if slot == slots.a: # avoid underflow at genesis
break
slot -= 1
false
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
proc getState(
dag: ChainDAGRef, bsi: BlockSlotId, state: var ForkedHashedBeaconState): bool =
## Load a state from the database given a block and a slot - this will first
## lookup the state root in the state root table then load the corresponding
## state, if it exists
if not dag.isStateCheckpoint(bsi):
return false
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let rollbackAddr =
# Any restore point will do as long as it's not the object being updated
if unsafeAddr(state) == unsafeAddr(dag.headState):
unsafeAddr dag.clearanceState
else:
unsafeAddr dag.headState
let v = addr state
func rollback() =
assign(v[], rollbackAddr[])
dag.db.getState(dag.cfg, bsi.bid.root, bsi.slot, state, rollback)
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
proc getStateByParent(
dag: ChainDAGRef, bid: BlockId, state: var ForkedHashedBeaconState): bool =
## Try to load the state referenced by the parent of the given `bid` - this
## state can be used to advance to the `bid` state itself.
var slot = bid.slot
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
let
summary = dag.db.getBeaconBlockSummary(bid.root).valueOr:
return false
parentMinSlot =
dag.db.getBeaconBlockSummary(summary.parent_root).
map(proc(x: auto): auto = x.slot).valueOr:
# in the cases that we don't have slot information, we'll search for the
# state for a few back from the `bid` slot - if there are gaps of empty
# slots larger than this, we will not be able to load the state using this
# trick
if slot.uint64 >= (EPOCHS_PER_STATE_SNAPSHOT * 2) * SLOTS_PER_EPOCH:
slot - (EPOCHS_PER_STATE_SNAPSHOT * 2) * SLOTS_PER_EPOCH
else:
Slot(0)
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
let rollbackAddr =
# Any restore point will do as long as it's not the object being updated
if unsafeAddr(state) == unsafeAddr(dag.headState):
unsafeAddr dag.clearanceState
else:
unsafeAddr dag.headState
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
let v = addr state
func rollback() =
assign(v[], rollbackAddr[])
dag.db.getState(
dag.cfg, summary.parent_root, parentMinSlot..slot, state, rollback)
proc currentSyncCommitteeForPeriod*(
dag: ChainDAGRef,
tmpState: var ForkedHashedBeaconState,
period: SyncCommitteePeriod): Opt[SyncCommittee] =
## Fetch a `SyncCommittee` for a given sync committee period.
## For non-finalized periods, follow the chain as selected by fork choice.
let lowSlot = max(dag.tail.slot, dag.cfg.ALTAIR_FORK_EPOCH.start_slot)
if period < lowSlot.sync_committee_period:
return err()
let
periodStartSlot = period.start_slot
syncCommitteeSlot = max(periodStartSlot, lowSlot)
bsi = ? dag.getBlockIdAtSlot(syncCommitteeSlot)
dag.withUpdatedState(tmpState, bsi) do:
withState(updatedState):
when stateFork >= BeaconStateFork.Altair:
ok forkyState.data.current_sync_committee
else: err()
do: err()
func isNextSyncCommitteeFinalized*(
dag: ChainDAGRef, period: SyncCommitteePeriod): bool =
let finalizedSlot = dag.finalizedHead.slot
if finalizedSlot < period.start_slot:
false
elif finalizedSlot < dag.cfg.ALTAIR_FORK_EPOCH.start_slot:
false # Fork epoch not necessarily tied to sync committee period boundary
else:
true
func firstNonFinalizedPeriod*(dag: ChainDAGRef): SyncCommitteePeriod =
if dag.finalizedHead.slot >= dag.cfg.ALTAIR_FORK_EPOCH.start_slot:
dag.finalizedHead.slot.sync_committee_period + 1
else:
dag.cfg.ALTAIR_FORK_EPOCH.sync_committee_period
proc updateBeaconMetrics(
state: ForkedHashedBeaconState, bid: BlockId, cache: var StateCache) =
# https://github.com/ethereum/beacon-metrics/blob/master/metrics.md#additional-metrics
# both non-negative, so difference can't overflow or underflow int64
beacon_head_root.set(bid.root.toGaugeValue)
beacon_head_slot.set(bid.slot.toGaugeValue)
withState(state):
beacon_pending_deposits.set(
(forkyState.data.eth1_data.deposit_count -
forkyState.data.eth1_deposit_index).toGaugeValue)
beacon_processed_deposits_total.set(
forkyState.data.eth1_deposit_index.toGaugeValue)
beacon_current_justified_epoch.set(
forkyState.data.current_justified_checkpoint.epoch.toGaugeValue)
beacon_current_justified_root.set(
forkyState.data.current_justified_checkpoint.root.toGaugeValue)
beacon_previous_justified_epoch.set(
forkyState.data.previous_justified_checkpoint.epoch.toGaugeValue)
beacon_previous_justified_root.set(
forkyState.data.previous_justified_checkpoint.root.toGaugeValue)
beacon_finalized_epoch.set(
forkyState.data.finalized_checkpoint.epoch.toGaugeValue)
beacon_finalized_root.set(
forkyState.data.finalized_checkpoint.root.toGaugeValue)
let active_validators = count_active_validators(
forkyState.data, forkyState.data.slot.epoch, cache).toGaugeValue
beacon_active_validators.set(active_validators)
beacon_current_active_validators.set(active_validators)
import blockchain_dag_light_client
export
blockchain_dag_light_client.getLightClientBootstrap,
blockchain_dag_light_client.getLightClientUpdateForPeriod,
blockchain_dag_light_client.getLightClientFinalityUpdate,
blockchain_dag_light_client.getLightClientOptimisticUpdate
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
proc putState(dag: ChainDAGRef, state: ForkedHashedBeaconState, bid: BlockId) =
# Store a state and its root
logScope:
blck = shortLog(bid)
stateSlot = shortLog(getStateField(state, slot))
stateRoot = shortLog(getStateRoot(state))
if not dag.isStateCheckpoint(BlockSlotId.init(bid, getStateField(state, slot))):
return
# Don't consider legacy tables here, they are slow to read so we'll want to
# rewrite things in the new table anyway.
if dag.db.containsState(getStateRoot(state), legacy = false):
return
let startTick = Moment.now()
# Ideally we would save the state and the root lookup cache in a single
# transaction to prevent database inconsistencies, but the state loading code
# is resilient against one or the other going missing
withState(state):
dag.db.putState(forkyState)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
debug "Stored state", putStateDur = Moment.now() - startTick
proc advanceSlots*(
dag: ChainDAGRef, state: var ForkedHashedBeaconState, slot: Slot, save: bool,
cache: var StateCache, info: var ForkedEpochInfo) =
# Given a state, advance it zero or more slots by applying empty slot
# processing - the state must be positioned at or before `slot`
doAssert getStateField(state, slot) <= slot
let stateBid = state.latest_block_id
while getStateField(state, slot) < slot:
let
preEpoch = getStateField(state, slot).epoch
loadStateCache(dag, cache, stateBid, getStateField(state, slot).epoch)
process_slots(
dag.cfg, state, getStateField(state, slot) + 1, cache, info,
dag.updateFlags).expect("process_slots shouldn't fail when state slot is correct")
if save:
dag.putState(state, stateBid)
# The reward information in the state transition is computed for epoch
# transitions - when transitioning into epoch N, the activities in epoch
# N-2 are translated into balance updates, and this is what we capture
# in the monitor. This may be inaccurate during a deep reorg (>1 epoch)
# which is an acceptable tradeoff for monitoring.
withState(state):
let postEpoch = forkyState.data.slot.epoch
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if preEpoch != postEpoch:
dag.validatorMonitor[].registerEpochInfo(
postEpoch, info, forkyState.data)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
proc applyBlock(
dag: ChainDAGRef, state: var ForkedHashedBeaconState, bid: BlockId,
cache: var StateCache, info: var ForkedEpochInfo): Result[void, cstring] =
loadStateCache(dag, cache, bid, getStateField(state, slot).epoch)
case dag.cfg.blockForkAtEpoch(bid.slot.epoch)
of BeaconBlockFork.Phase0:
let data = getBlock(dag, bid, phase0.TrustedSignedBeaconBlock).valueOr:
return err("Block load failed")
state_transition(
dag.cfg, state, data, cache, info,
dag.updateFlags + {slotProcessed}, noRollback)
of BeaconBlockFork.Altair:
let data = getBlock(dag, bid, altair.TrustedSignedBeaconBlock).valueOr:
return err("Block load failed")
state_transition(
dag.cfg, state, data, cache, info,
dag.updateFlags + {slotProcessed}, noRollback)
of BeaconBlockFork.Bellatrix:
let data = getBlock(dag, bid, bellatrix.TrustedSignedBeaconBlock).valueOr:
return err("Block load failed")
state_transition(
dag.cfg, state, data, cache, info,
dag.updateFlags + {slotProcessed}, noRollback)
of BeaconBlockFork.Capella:
let data = getBlock(dag, bid, capella.TrustedSignedBeaconBlock).valueOr:
return err("Block load failed")
state_transition(
dag.cfg, state, data, cache, info,
dag.updateFlags + {slotProcessed}, noRollback)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
proc init*(T: type ChainDAGRef, cfg: RuntimeConfig, db: BeaconChainDB,
validatorMonitor: ref ValidatorMonitor, updateFlags: UpdateFlags,
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
eraPath = ".",
onBlockCb: OnBlockCallback = nil, onHeadCb: OnHeadCallback = nil,
onReorgCb: OnReorgCallback = nil, onFinCb: OnFinalizedCallback = nil,
vanityLogs = default(VanityLogs),
lcDataConfig = default(LightClientDataConfig)): ChainDAGRef =
cfg.checkForkConsistency()
doAssert updateFlags - {
strictVerification, enableTestFeatures, lowParticipation
} == {}, "Other flags not supported in ChainDAG"
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# TODO we require that the db contains both a head and a tail block -
# asserting here doesn't seem like the right way to go about it however..
# Tail is the first block for which we can construct a state - either
# genesis or a checkpoint
let
startTick = Moment.now()
genesisRoot = db.getGenesisBlock()
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
tailRoot = db.getTailBlock().expect(
"preInit should have initialized the database with a tail block root")
tail = db.getBlockId(tailRoot).expect(
"tail block summary in database, database corrupt?")
headRoot = db.getHeadBlock().expect("head root, database corrupt?")
head = db.getBlockId(headRoot).expect("head block id, database corrupt?")
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# Have to be careful with this instance, it is not yet fully initialized so
# as to avoid having to allocate a separate "init" state
dag = ChainDAGRef(
db: db,
validatorMonitor: validatorMonitor,
genesis: genesisRoot.map(
proc(x: auto): auto = BlockId(root: x, slot: GENESIS_SLOT)),
tail: tail,
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# The only allowed flag right now is strictVerification, as the others all
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# allow skipping some validation.
updateFlags: updateFlags * {
strictVerification, enableTestFeatures, lowParticipation
},
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
cfg: cfg,
vanityLogs: vanityLogs,
lcDataStore: initLightClientDataStore(
lcDataConfig, cfg, db.getLightClientDataDB()),
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
onBlockAdded: onBlockCb,
onHeadChanged: onHeadCb,
onReorgHappened: onReorgCb,
onFinHappened: onFinCb
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
)
loadTick = Moment.now()
2020-10-06 15:32:17 +00:00
var
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
headRef, curRef: BlockRef
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
# When starting from a checkpoint with an empty block, we'll store the state
# "ahead" of the head slot - this slot would be considered finalized
slot = max(head.slot, (tail.slot.epoch + 1).start_slot)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# To know the finalized checkpoint of the head, we need to recreate its
# state - the tail is implicitly finalized, and if we have a finalized block
# table, that provides another hint
finalizedSlot = db.finalizedBlocks.high.get(tail.slot)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
newFinalized: seq[BlockId]
cache: StateCache
foundHeadState = false
headBlocks: seq[BlockRef]
# Load head -> finalized, or all summaries in case the finalized block table
# hasn't been written yet
for blck in db.getAncestorSummaries(head.root):
# The execution block root gets filled in as needed
let newRef = BlockRef.init(blck.root, none Eth2Digest, blck.summary.slot)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if headRef == nil:
headRef = newRef
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if curRef != nil:
link(newRef, curRef)
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
curRef = newRef
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.forkBlocks.incl(KeyedBlockRef.init(curRef))
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if not foundHeadState:
foundHeadState = db.getState(
cfg, blck.root, blck.summary.slot..slot, dag.headState, noRollback)
slot = blck.summary.slot
chaindag: don't keep backfill block table in memory (#3429) This PR names and documents the concept of the archive: a range of slots for which we have degraded functionality in terms of historical access - in particular: * we don't support rewinding to states in this range * we don't keep an in-memory representation of the block dag The archive de-facto exists in a trusted-node-synced node, but this PR gives it a name and drops the in-memory digest index. In order to satisfy `GetBlocksByRange` requests, we ensure that we have blocks for the entire archive period via backfill. Future versions may relax this further, adding a "pre-archive" period that is fully pruned. During by-slot searches in the archive (both for libp2p and rest requests), an extra database lookup is used to covert the given `slot` to a `root` - future versions will avoid this using era files which natively are indexed by `slot`. That said, the lookup is quite fast compared to the actual block loading given how trivial the table is - it's hard to measure, even. A collateral benefit of this PR is that checkpoint-synced nodes will see 100-200MB memory usage savings, thanks to the dropped in-memory cache - future pruning work will bring this benefit to full nodes as well. * document chaindag storage architecture and assumptions * look up parent using block id instead of full block in clearance (future-proofing the code against a future in which blocks come from era files) * simplify finalized block init, always writing the backfill portion to db at startup (to ensure lookups work as expected) * preallocate some extra memory for finalized blocks, to avoid immediate realloc
2022-02-26 18:16:19 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if not foundHeadState:
# When the database has been written with a pre-fork version of the
# software, it may happen that blocks produced using an "unforked"
# chain get written to the database - we need to skip such blocks
# when loading the database with a fork-compatible version
if containsBlock(cfg, db, curRef.slot, curRef.root):
headBlocks.add curRef
else:
if headBlocks.len > 0:
fatal "Missing block needed to create head state, database corrupt?",
curRef = shortLog(curRef)
quit 1
# Without the block data we can't form a state for this root, so
# we'll need to move the head back
headRef = nil
dag.forkBlocks.excl(KeyedBlockRef.init(curRef))
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
if curRef.slot <= finalizedSlot:
# Only non-finalized slots get a `BlockRef`
break
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
let summariesTick = Moment.now()
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if not foundHeadState:
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
if not dag.getStateByParent(curRef.bid, dag.headState):
fatal "Could not load head state, database corrupt?",
head = shortLog(head), tail = shortLog(dag.tail)
quit 1
block:
# EpochRef needs an epoch boundary state
assign(dag.epochRefState, dag.headState)
var info: ForkedEpochInfo
while headBlocks.len > 0:
dag.applyBlock(
dag.headState, headBlocks.pop().bid, cache,
info).expect("head blocks should apply")
dag.head = headRef
assign(dag.clearanceState, dag.headState)
if dag.headState.latest_block_root == tail.root:
# In case we started from a checkpoint with an empty slot
finalizedSlot = getStateField(dag.headState, slot)
finalizedSlot =
max(
finalizedSlot,
getStateField(dag.headState, finalized_checkpoint).epoch.start_slot)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let
configFork = case dag.headState.kind
of BeaconStateFork.Phase0: genesisFork(cfg)
of BeaconStateFork.Altair: altairFork(cfg)
of BeaconStateFork.Bellatrix: bellatrixFork(cfg)
of BeaconStateFork.Capella: capellaFork(cfg)
stateFork = getStateField(dag.headState, fork)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if stateFork != configFork:
error "State from database does not match network, check --network parameter",
tail = dag.tail, headRef, stateFork, configFork
quit 1
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
# Need to load state to find genesis validators root, before loading era db
dag.era = EraDB.new(
cfg, eraPath, getStateField(dag.headState, genesis_validators_root))
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# We used an interim finalizedHead while loading the head state above - now
# that we have loaded the dag up to the finalized slot, we can also set
# finalizedHead to its real value
dag.finalizedHead = headRef.atSlot(finalizedSlot)
dag.lastPrunePoint = dag.finalizedHead.toBlockSlotId().expect("not nil")
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.heads = @[headRef]
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
doAssert dag.finalizedHead.blck != nil,
"The finalized head should exist at the slot"
doAssert dag.finalizedHead.blck.parent == nil,
"...but that's the last BlockRef with a parent"
block: # Top up finalized blocks
if db.finalizedBlocks.high.isNone or
db.finalizedBlocks.high.get() < dag.finalizedHead.blck.slot:
info "Loading finalized blocks",
finHigh = db.finalizedBlocks.high,
finalizedHead = shortLog(dag.finalizedHead)
for blck in db.getAncestorSummaries(dag.finalizedHead.blck.root):
if db.finalizedBlocks.high.isSome and
blck.summary.slot <= db.finalizedBlocks.high.get:
break
# Versions prior to 1.7.0 did not store finalized blocks in the
# database, and / or the application might have crashed between the head
# and finalized blocks updates.
newFinalized.add(BlockId(slot: blck.summary.slot, root: blck.root))
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let finalizedBlocksTick = Moment.now()
db.updateFinalizedBlocks(newFinalized)
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
block:
let finalized = db.finalizedBlocks.get(db.finalizedBlocks.high.get()).expect(
"tail at least")
if finalized != dag.finalizedHead.blck.root:
error "Head does not lead to finalized block, database corrupt?",
head = shortLog(head), finalizedHead = shortLog(dag.finalizedHead),
tail = shortLog(dag.tail), finalized = shortLog(finalized)
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
quit 1
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.backfill = block:
let backfillSlot = db.finalizedBlocks.low.expect("tail at least")
if backfillSlot < dag.tail.slot:
let backfillRoot = db.finalizedBlocks.get(backfillSlot).expect(
"low to be loadable")
db.getBeaconBlockSummary(backfillRoot).expect(
"Backfill block must have a summary: " & $backfillRoot)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
else:
db.getBeaconBlockSummary(dag.tail.root).expect(
"Tail block must have a summary: " & $dag.tail.root)
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
dag.forkDigests = newClone ForkDigests.init(
cfg, getStateField(dag.headState, genesis_validators_root))
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
withState(dag.headState):
dag.validatorMonitor[].registerState(forkyState.data)
updateBeaconMetrics(dag.headState, dag.head.bid, cache)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let finalizedTick = Moment.now()
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
if dag.backfill.slot > 0: # See if we can frontfill blocks from era files
dag.frontfillBlocks = newSeqOfCap[Eth2Digest](dag.backfill.slot.int)
let
historical_roots = getStateField(dag.headState, historical_roots).asSeq()
var
blocks = 0
# Here, we'll build up the slot->root mapping in memory for the range of
# blocks from genesis to backfill, if possible.
for bid in dag.era.getBlockIds(historical_roots, Slot(0), Eth2Digest()):
if bid.slot >= dag.backfill.slot:
# If we end up in here, we failed the root comparison just below in
# an earlier iteration
fatal "Era summaries don't lead up to backfill, database or era files corrupt?",
bid
quit 1
# In BeaconState.block_roots, empty slots are filled with the root of
# the previous block - in our data structure, we use a zero hash instead
dag.frontfillBlocks.setLen(bid.slot.int + 1)
dag.frontfillBlocks[bid.slot.int] = bid.root
if bid.root == dag.backfill.parent_root:
# We've reached the backfill point, meaning blocks are available
# in the sqlite database from here onwards - remember this point in
# time so that we can write summaries to the database - it's a lot
# faster to load from database than to iterate over era files with
# the current naive era file reader.
reset(dag.backfill)
dag.updateFrontfillBlocks()
break
blocks += 1
if blocks > 0:
info "Front-filled blocks from era files", blocks
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
let frontfillTick = Moment.now()
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
# Fill validator key cache in case we're loading an old database that doesn't
# have a cache
dag.updateValidatorKeys(getStateField(dag.headState, validators).asSeq())
withState(dag.headState):
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
when stateFork >= BeaconStateFork.Altair:
dag.headSyncCommittees = forkyState.data.get_sync_committee_cache(cache)
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
info "Block DAG initialized",
head = shortLog(dag.head),
finalizedHead = shortLog(dag.finalizedHead),
tail = shortLog(dag.tail),
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
backfill = (dag.backfill.slot, shortLog(dag.backfill.parent_root)),
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
loadDur = loadTick - startTick,
summariesDur = summariesTick - loadTick,
finalizedDur = finalizedTick - summariesTick,
era: load blocks and states (#3394) * era: load blocks and states Era files contain finalized history and can be thought of as an alternative source for block and state data that allows clients to avoid syncing this information from the P2P network - the P2P network is then used to "top up" the client with the most recent data. They can be freely shared in the community via whatever means (http, torrent, etc) and serve as a permanent cold store of consensus data (and, after the merge, execution data) for history buffs and bean counters alike. This PR gently introduces support for loading blocks and states in two cases: block requests from rest/p2p and frontfilling when doing checkpoint sync. The era files are used as a secondary source if the information is not found in the database - compared to the database, there are a few key differences: * the database stores the block indexed by block root while the era file indexes by slot - the former is used only in rest, while the latter is used both by p2p and rest. * when loading blocks from era files, the root is no longer trivially available - if it is needed, it must either be computed (slow) or cached (messy) - the good news is that for p2p requests, it is not needed * in era files, "framed" snappy encoding is used while in the database we store unframed snappy - for p2p2 requests, the latter requires recompression while the former could avoid it * front-filling is the process of using era files to replace backfilling - in theory this front-filling could happen from any block and front-fills with gaps could also be entertained, but our backfilling algorithm cannot take advantage of this because there's no (simple) way to tell it to "skip" a range. * front-filling, as implemented, is a bit slow (10s to load mainnet): we load the full BeaconState for every era to grab the roots of the blocks - it would be better to partially load the state - as such, it would also be good to be able to partially decompress snappy blobs * lookups from REST via root are served by first looking up a block summary in the database, then using the slot to load the block data from the era file - however, there needs to be an option to create the summary table from era files to fully support historical queries To test this, `ncli_db` has an era file exporter: the files it creates should be placed in an `era` folder next to `db` in the data directory. What's interesting in particular about this setup is that `db` remains as the source of truth for security purposes - it stores the latest synced head root which in turn determines where a node "starts" its consensus participation - the era directory however can be freely shared between nodes / people without any (significant) security implications, assuming the era files are consistent / not broken. There's lots of future improvements to be had: * we can drop the in-memory `BlockRef` index almost entirely - at this point, resident memory usage of Nimbus should drop to a cool 500-600 mb * we could serve era files via REST trivially: this would drop backfill times to whatever time it takes to download the files - unlike the current implementation that downloads block by block, downloading an era at a time almost entirely cuts out request overhead * we can "reasonably" recreate detailed state history from almost any point in time, turning an O(slot) process into O(1) effectively - we'll still need caches and indices to do this with sufficient efficiency for the rest api, but at least it cuts the whole process down to minutes instead of hours, for arbitrary points in time * CI: ignore failures with Nim-1.6 (temporary) * test fixes Co-authored-by: Ștefan Talpalaru <stefantalpalaru@yahoo.com>
2022-03-23 08:58:17 +00:00
frontfillDur = frontfillTick - finalizedTick,
keysDur = Moment.now() - frontfillTick
dag.initLightClientDataCache()
# If these aren't actually optimistic, the first fcU will resolve that
withState(dag.headState):
when stateFork >= BeaconStateFork.Bellatrix:
template executionPayloadHeader(): auto =
forkyState().data.latest_execution_payload_header
const emptyExecutionPayloadHeader =
default(type(executionPayloadHeader))
if executionPayloadHeader != emptyExecutionPayloadHeader:
dag.optimisticRoots.incl dag.head.root
dag.optimisticRoots.incl dag.finalizedHead.blck.root
dag
template genesis_validators_root*(dag: ChainDAGRef): Eth2Digest =
getStateField(dag.headState, genesis_validators_root)
proc genesisBlockRoot*(dag: ChainDAGRef): Eth2Digest =
dag.db.getGenesisBlock().expect("DB must be initialized with genesis block")
func getEpochRef*(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag: ChainDAGRef, state: ForkedHashedBeaconState, cache: var StateCache): EpochRef =
## Get a cached `EpochRef` or construct one based on the given state - always
## returns an EpochRef instance
let
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
bid = state.latest_block_id
epoch = state.get_current_epoch()
dag.findEpochRef(bid, epoch).valueOr:
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let res = EpochRef.init(dag, state, cache)
dag.putEpochRef(res)
res
proc getEpochRef*(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag: ChainDAGRef, bid: BlockId, epoch: Epoch,
preFinalized: bool): Result[EpochRef, cstring] =
## Return a cached EpochRef or construct one from the database, if possible -
## returns `none` on failure.
##
## When `preFinalized` is true, include epochs from before the finalized
## checkpoint in the search - this potentially can result in long processing
## times due to state replays.
##
## Requests for epochs >= dag.finalizedHead.slot.epoch always return an
## instance. One must be careful to avoid race conditions in `async` code
## where the finalized head might change during an `await`.
##
## Requests for epochs < dag.finalizedHead.slot.epoch may fail, either because
## the search was limited by the `preFinalized` flag or because state history
## has been pruned - `none` will be returned in this case.
if not preFinalized and epoch < dag.finalizedHead.slot.epoch:
return err("Requesting pre-finalized EpochRef")
if bid.slot < dag.tail.slot or epoch < dag.tail.slot.epoch:
return err("Requesting EpochRef for pruned state")
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let epochRef = dag.findEpochRef(bid, epoch)
if epochRef.isOk():
beacon_state_data_cache_hits.inc
return ok epochRef.get()
beacon_state_data_cache_misses.inc
let
ancestor = dag.epochAncestor(bid, epoch).valueOr:
# If we got in here, the bid must be unknown or we would have gotten
# _some_ ancestor (like the tail)
return err("Requesting EpochRef for non-canonical block")
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
var cache: StateCache
if not updateState(dag, dag.epochRefState, ancestor, false, cache):
return err("Could not load requested state")
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
ok(dag.getEpochRef(dag.epochRefState, cache))
proc getEpochRef*(
dag: ChainDAGRef, blck: BlockRef, epoch: Epoch,
preFinalized: bool): Result[EpochRef, cstring] =
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.getEpochRef(blck.bid, epoch, preFinalized)
proc getFinalizedEpochRef*(dag: ChainDAGRef): EpochRef =
dag.getEpochRef(
dag.finalizedHead.blck, dag.finalizedHead.slot.epoch, false).expect(
"getEpochRef for finalized head should always succeed")
proc getShufflingRef*(
dag: ChainDAGRef, blck: BlockRef, epoch: Epoch,
preFinalized: bool): Opt[ShufflingRef] =
## Return the shuffling in the given history and epoch - this potentially is
## faster than returning a full EpochRef because the shuffling is determined
## an epoch in advance and therefore is less sensitive to reorgs
let shufflingRef = dag.findShufflingRef(blck.bid, epoch)
if shufflingRef.isNone:
# TODO here, we could check the existing cached states and see if any one
# has the right dependent root - unlike EpochRef, we don't need an _exact_
# epoch match
let epochRef = dag.getEpochRef(blck, epoch, preFinalized).valueOr:
return Opt.none ShufflingRef
dag.putShufflingRef(epochRef.shufflingRef)
Opt.some epochRef.shufflingRef
else:
shufflingRef
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
func stateCheckpoint*(dag: ChainDAGRef, bsi: BlockSlotId): BlockSlotId =
## The first ancestor BlockSlot that is a state checkpoint
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
var bsi = bsi
while not dag.isStateCheckpoint(bsi):
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if bsi.isProposed:
bsi.bid = dag.parent(bsi.bid).valueOr:
break
else:
bsi.slot = bsi.slot - 1
bsi
template forkAtEpoch*(dag: ChainDAGRef, epoch: Epoch): Fork =
forkAtEpoch(dag.cfg, epoch)
proc getBlockRange*(
dag: ChainDAGRef, startSlot: Slot, skipStep: uint64,
output: var openArray[BlockId]): Natural =
## This function populates an `output` buffer of blocks
## with a slots ranging from `startSlot` up to, but not including,
## `startSlot + skipStep * output.len`, skipping any slots that don't have
## a block.
##
## Blocks will be written to `output` from the end without gaps, even if
## a block is missing in a particular slot. The return value shows how
## many slots were missing blocks - to iterate over the result, start
## at this index.
##
## If there were no blocks in the range, `output.len` will be returned.
let
requestedCount = output.lenu64
headSlot = dag.head.slot
trace "getBlockRange entered",
head = shortLog(dag.head.root), requestedCount, startSlot, skipStep, headSlot
if startSlot < dag.backfill.slot:
notice "Got request for pre-backfill slot",
startSlot, backfillSlot = dag.backfill.slot
return output.len
if headSlot <= startSlot or requestedCount == 0:
return output.len # Identical to returning an empty set of block as indicated above
let
runway = uint64(headSlot - startSlot)
# This is the number of blocks that will follow the start block
extraSlots = min(runway div skipStep, requestedCount - 1)
# If `skipStep` is very large, `extraSlots` should be 0 from
# the previous line, so `endSlot` will be equal to `startSlot`:
endSlot = startSlot + extraSlots * skipStep
var
curSlot = endSlot
o = output.len
# Process all blocks that follow the start block (may be zero blocks)
while curSlot > startSlot:
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
let bs = dag.getBlockIdAtSlot(curSlot)
if bs.isSome and bs.get().isProposed():
o -= 1
output[o] = bs.get().bid
curSlot -= skipStep
# Handle start slot separately (to avoid underflow when computing curSlot)
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
let bs = dag.getBlockIdAtSlot(startSlot)
if bs.isSome and bs.get().isProposed():
o -= 1
output[o] = bs.get().bid
o # Return the index of the first non-nil item in the output
proc updateState*(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag: ChainDAGRef, state: var ForkedHashedBeaconState, bsi: BlockSlotId,
save: bool, cache: var StateCache): bool =
## Rewind or advance state such that it matches the given block and slot -
## this may include replaying from an earlier snapshot if blck is on a
## different branch or has advanced to a higher slot number than slot
## If `bs.slot` is higher than `bs.blck.slot`, `updateState` will fill in
## with empty/non-block slots
# First, see if we're already at the requested block. If we are, also check
# that the state has not been advanced past the desired block - if it has,
# an earlier state must be loaded since there's no way to undo the slot
# transitions
let
startTick = Moment.now()
current {.used.} = withState(state):
BlockSlotId.init(forkyState.latest_block_id, forkyState.data.slot)
var
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
ancestors: seq[BlockId]
found = false
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
template exactMatch(state: ForkedHashedBeaconState, bsi: BlockSlotId): bool =
# The block is the same and we're at an early enough slot - the state can
# be used to arrive at the desired blockslot
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
state.matches_block_slot(bsi.bid.root, bsi.slot)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
template canAdvance(state: ForkedHashedBeaconState, bsi: BlockSlotId): bool =
# The block is the same and we're at an early enough slot - the state can
# be used to arrive at the desired blockslot
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
state.can_advance_slots(bsi.bid.root, bsi.slot)
# Fast path: check all caches for an exact match - this is faster than
# advancing a state where there's epoch processing to do, by a wide margin -
# it also avoids `hash_tree_root` for slot processing
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if exactMatch(state, bsi):
found = true
elif not save:
# When required to save states, we cannot rely on the caches because that
# would skip the extra processing that save does - not all information that
# goes into the database is cached
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if exactMatch(dag.headState, bsi):
assign(state, dag.headState)
found = true
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
elif exactMatch(dag.clearanceState, bsi):
assign(state, dag.clearanceState)
found = true
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
elif exactMatch(dag.epochRefState, bsi):
assign(state, dag.epochRefState)
found = true
const RewindBlockThreshold = 64
if not found:
# No exact match found - see if any in-memory state can be used as a base
# onto which we can apply a few blocks - there's a tradeoff here between
# loading the state from disk and performing the block applications
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
var cur = bsi
while ancestors.len < RewindBlockThreshold:
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if isZero(cur.bid.root): # tail reached
break
if canAdvance(state, cur): # Typical case / fast path when there's no reorg
found = true
break
if not save: # see above
if canAdvance(dag.headState, cur):
assign(state, dag.headState)
found = true
break
if canAdvance(dag.clearanceState, cur):
assign(state, dag.clearanceState)
found = true
break
if canAdvance(dag.epochRefState, cur):
assign(state, dag.epochRefState)
found = true
break
if cur.isProposed():
# This is not an empty slot, so the block will need to be applied to
# eventually reach bs
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
ancestors.add(cur.bid)
# Move slot by slot to capture epoch boundary states
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# TODO https://github.com/nim-lang/Nim/issues/19613
cur = dag.parentOrSlot(cur).valueOr:
break
if not found:
debug "UpdateStateData cache miss",
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
current = shortLog(current), target = shortLog(bsi)
# Either the state is too new or was created by applying a different block.
# We'll now resort to loading the state from the database then reapplying
# blocks until we reach the desired point in time.
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
var cur = bsi
ancestors.setLen(0)
# Look for a state in the database and load it - as long as it cannot be
# found, keep track of the blocks that are needed to reach it from the
# state that eventually will be found.
# If we hit the tail, it means that we've reached a point for which we can
# no longer recreate history - this happens for example when starting from
# a checkpoint block
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let startEpoch = bsi.slot.epoch
while not canAdvance(state, cur) and
not dag.db.getState(dag.cfg, cur.bid.root, cur.slot, state, noRollback):
# There's no state saved for this particular BlockSlot combination, and
# the state we have can't trivially be advanced (in case it was older than
# RewindBlockThreshold), keep looking..
if cur.isProposed():
# This is not an empty slot, so the block will need to be applied to
# eventually reach bs
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
ancestors.add(cur.bid)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if cur.slot == GENESIS_SLOT or
2022-04-10 20:10:22 +00:00
(cur.slot.epoch + uint64(EPOCHS_PER_STATE_SNAPSHOT) * 2 < startEpoch):
# We've either walked two full state snapshot lengths or hit the tail
# and still can't find a matching state: this can happen when
# starting the node from an arbitrary finalized checkpoint and not
# backfilling the states
notice "Request for pruned historical state",
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
request = shortLog(bsi), tail = shortLog(dag.tail),
cur = shortLog(cur)
return false
# Move slot by slot to capture epoch boundary states
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# TODO https://github.com/nim-lang/Nim/issues/19613
cur = dag.parentOrSlot(cur).valueOr:
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
if not dag.getStateByParent(cur.bid, state):
notice "Request for pruned historical state",
request = shortLog(bsi), tail = shortLog(dag.tail),
cur = shortLog(cur)
return false
break
beacon_state_rewinds.inc()
# Starting state has been assigned, either from memory or database
let
assignTick = Moment.now()
ancestor {.used.} = withState(state):
BlockSlotId.init(forkyState.latest_block_id, forkyState.data.slot)
ancestorRoot {.used.} = getStateRoot(state)
var info: ForkedEpochInfo
# Time to replay all the blocks between then and now
for i in countdown(ancestors.len - 1, 0):
# Because the ancestors are in the database, there's no need to persist them
# again. Also, because we're applying blocks that were loaded from the
# database, we can skip certain checks that have already been performed
# before adding the block to the database.
if (let res = dag.applyBlock(state, ancestors[i], cache, info); res.isErr):
warn "Failed to apply block from database",
blck = shortLog(ancestors[i]),
state_bid = shortLog(state.latest_block_id),
error = res.error()
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
return false
# ...and make sure to process empty slots as requested
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.advanceSlots(state, bsi.slot, save, cache, info)
# ...and make sure to load the state cache, if it exists
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
loadStateCache(dag, cache, bsi.bid, getStateField(state, slot).epoch)
let
assignDur = assignTick - startTick
replayDur = Moment.now() - assignTick
# TODO https://github.com/status-im/nim-chronicles/issues/108
if (assignDur + replayDur) >= 250.millis:
# This might indicate there's a cache that's not in order or a disk that is
# too slow - for now, it's here for investigative purposes and the cutoff
# time might need tuning
info "State replayed",
blocks = ancestors.len,
slots = getStateField(state, slot) - ancestor.slot,
current = shortLog(current),
ancestor = shortLog(ancestor),
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
target = shortLog(bsi),
ancestorStateRoot = shortLog(ancestorRoot),
targetStateRoot = shortLog(getStateRoot(state)),
found,
assignDur,
replayDur
elif ancestors.len > 0:
debug "State replayed",
blocks = ancestors.len,
slots = getStateField(state, slot) - ancestor.slot,
current = shortLog(current),
ancestor = shortLog(ancestor),
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
target = shortLog(bsi),
ancestorStateRoot = shortLog(ancestorRoot),
targetStateRoot = shortLog(getStateRoot(state)),
found,
assignDur,
replayDur
else: # Normal case!
trace "State advanced",
blocks = ancestors.len,
slots = getStateField(state, slot) - ancestor.slot,
current = shortLog(current),
ancestor = shortLog(ancestor),
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
target = shortLog(bsi),
ancestorStateRoot = shortLog(ancestorRoot),
targetStateRoot = shortLog(getStateRoot(state)),
found,
assignDur,
replayDur
true
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
proc delState(dag: ChainDAGRef, bsi: BlockSlotId) =
# Delete state state and mapping for a particular block+slot
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if not dag.isStateCheckpoint(bsi):
return # We only ever save epoch states
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if (let root = dag.db.getStateRoot(bsi.bid.root, bsi.slot); root.isSome()):
dag.db.delState(root.get())
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.db.delStateRoot(bsi.bid.root, bsi.slot)
proc pruneBlockSlot(dag: ChainDAGRef, bs: BlockSlot) =
# TODO: should we move that disk I/O to `onSlotEnd`
dag.delState(bs.toBlockSlotId().expect("not nil"))
if bs.isProposed():
# Update light client data
dag.deleteLightClientData(bs.blck.bid)
dag.optimisticRoots.excl bs.blck.root
dag.forkBlocks.excl(KeyedBlockRef.init(bs.blck))
dag.db.delBlock(bs.blck.root)
proc pruneBlocksDAG(dag: ChainDAGRef) =
## This prunes the block DAG
## This does NOT prune the cached state checkpoints and EpochRef
## This must be done after a new finalization point is reached
## to invalidate pending blocks or attestations referring
## to a now invalid fork.
##
## This does NOT update the `dag.lastPrunePoint` field.
## as the caches and fork choice can be pruned at a later time.
# Clean up block refs, walking block by block
let startTick = Moment.now()
# Finalization means that we choose a single chain as the canonical one -
# it also means we're no longer interested in any branches from that chain
# up to the finalization point
let hlen = dag.heads.len
for i in 0..<hlen:
let n = hlen - i - 1
let head = dag.heads[n]
if dag.finalizedHead.blck.isAncestorOf(head):
continue
var cur = head.atSlot()
# The block whose parent is nil is the `BlockRef` that's part of the
# canonical chain but has now been finalized - in theory there could be
# states at empty slot iff the fork had epoch-long gaps where the epoch
# transition was not on the canonical chain - these will not properly get
# cleaned up by the current logic - but they should also be rare
# TODO clean up the above as well
doAssert dag.finalizedHead.blck.parent == nil,
"finalizedHead parent should have been pruned from memory already"
while cur.blck.parent != nil:
dag.pruneBlockSlot(cur)
cur = cur.parentOrSlot
dag.heads.del(n)
debug "Pruned the blockchain DAG",
currentCandidateHeads = dag.heads.len,
prunedHeads = hlen - dag.heads.len,
dagPruneDur = Moment.now() - startTick
# https://github.com/ethereum/consensus-specs/blob/v1.3.0-alpha.1/sync/optimistic.md#helpers
template is_optimistic*(dag: ChainDAGRef, root: Eth2Digest): bool =
root in dag.optimisticRoots
proc markBlockVerified*(
dag: ChainDAGRef, quarantine: var Quarantine, root: Eth2Digest) =
# Might be called when block was not optimistic to begin with, or had been
# but already had been marked verified.
if not dag.is_optimistic(root):
return
var cur = dag.getBlockRef(root).valueOr:
return
logScope: blck = shortLog(cur)
debug "markBlockVerified"
while true:
if not dag.is_optimistic(cur.bid.root):
return
dag.optimisticRoots.excl cur.bid.root
debug "markBlockVerified ancestor"
if cur.parent.isNil:
break
cur = cur.parent
iterator syncSubcommittee*(
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
syncCommittee: openArray[ValidatorIndex],
subcommitteeIdx: SyncSubcommitteeIndex): ValidatorIndex =
var i = subcommitteeIdx.asInt * SYNC_SUBCOMMITTEE_SIZE
let onePastEndIdx = min(syncCommittee.len, i + SYNC_SUBCOMMITTEE_SIZE)
while i < onePastEndIdx:
yield syncCommittee[i]
inc i
iterator syncSubcommitteePairs*(
syncCommittee: openArray[ValidatorIndex],
subcommitteeIdx: SyncSubcommitteeIndex): tuple[validatorIdx: ValidatorIndex,
subcommitteeIdx: int] =
var i = subcommitteeIdx.asInt * SYNC_SUBCOMMITTEE_SIZE
let onePastEndIdx = min(syncCommittee.len, i + SYNC_SUBCOMMITTEE_SIZE)
while i < onePastEndIdx:
yield (syncCommittee[i], i)
inc i
func syncCommitteeParticipants*(dag: ChainDAGRef,
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
slot: Slot): seq[ValidatorIndex] =
withState(dag.headState):
when stateFork >= BeaconStateFork.Altair:
let
period = sync_committee_period(slot)
curPeriod = sync_committee_period(forkyState.data.slot)
if period == curPeriod:
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
@(dag.headSyncCommittees.current_sync_committee)
elif period == curPeriod + 1:
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
@(dag.headSyncCommittees.next_sync_committee)
else: @[]
else:
@[]
func getSubcommitteePositionsAux(
dag: ChainDAGRef,
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
syncCommittee: openArray[ValidatorIndex],
subcommitteeIdx: SyncSubcommitteeIndex,
validatorIdx: uint64): seq[uint64] =
var pos = 0'u64
for valIdx in syncCommittee.syncSubcommittee(subcommitteeIdx):
if validatorIdx == uint64(valIdx):
result.add pos
inc pos
func getSubcommitteePositions*(
dag: ChainDAGRef,
slot: Slot,
subcommitteeIdx: SyncSubcommitteeIndex,
validatorIdx: uint64): seq[uint64] =
withState(dag.headState):
when stateFork >= BeaconStateFork.Altair:
let
period = sync_committee_period(slot)
curPeriod = sync_committee_period(forkyState.data.slot)
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
template search(syncCommittee: openArray[ValidatorIndex]): seq[uint64] =
dag.getSubcommitteePositionsAux(
syncCommittee, subcommitteeIdx, validatorIdx)
if period == curPeriod:
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
search(dag.headSyncCommittees.current_sync_committee)
elif period == curPeriod + 1:
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
search(dag.headSyncCommittees.next_sync_committee)
else: @[]
else:
@[]
template syncCommitteeParticipants*(
dag: ChainDAGRef,
slot: Slot,
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
subcommitteeIdx: SyncSubcommitteeIndex): seq[ValidatorIndex] =
toSeq(syncSubcommittee(dag.syncCommitteeParticipants(slot), subcommitteeIdx))
iterator syncCommitteeParticipants*(
dag: ChainDAGRef,
slot: Slot,
subcommitteeIdx: SyncSubcommitteeIndex,
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
aggregationBits: SyncCommitteeAggregationBits): ValidatorIndex =
for pos, valIdx in dag.syncCommitteeParticipants(slot, subcommitteeIdx):
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
if pos < aggregationBits.bits and aggregationBits[pos]:
yield valIdx
func needStateCachesAndForkChoicePruning*(dag: ChainDAGRef): bool =
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.lastPrunePoint != dag.finalizedHead.toBlockSlotId().expect("not nil")
proc pruneStateCachesDAG*(dag: ChainDAGRef) =
## This prunes the cached state checkpoints and EpochRef
## This does NOT prune the state associated with invalidated blocks on a fork
## They are pruned via `pruneBlocksDAG`
##
## This updates the `dag.lastPrunePoint` variable
doAssert dag.needStateCachesAndForkChoicePruning()
let startTick = Moment.now()
block: # Remove states, walking slot by slot
# We remove all state checkpoints that come _before_ the current finalized
# head, as we might frequently be asked to replay states from the
# finalized checkpoint and onwards (for example when validating blocks and
# attestations)
var
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
finPoint = dag.finalizedHead.toBlockSlotId().expect("not nil")
cur = dag.parentOrSlot(dag.stateCheckpoint(finPoint))
prev = dag.parentOrSlot(dag.stateCheckpoint(dag.lastPrunePoint))
while cur.isSome and prev.isSome and cur.get() != prev.get():
if not isFinalizedStateSnapshot(cur.get().slot) and
cur.get().slot != dag.tail.slot:
dag.delState(cur.get())
# TODO https://github.com/nim-lang/Nim/issues/19613
let tmp = cur.get()
cur = dag.parentOrSlot(tmp)
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let statePruneTick = Moment.now()
block: # Clean up old EpochRef instances
# After finalization, we can clear up the epoch cache and save memory -
# it will be recomputed if needed
dag.epochRefs.delIt(it.epoch < dag.finalizedHead.slot.epoch)
dag.shufflingRefs.delIt(it.epoch < dag.finalizedHead.slot.epoch)
let epochRefPruneTick = Moment.now()
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.lastPrunePoint = dag.finalizedHead.toBlockSlotId().expect("not nil")
2021-03-17 06:30:16 +00:00
debug "Pruned the state checkpoints and DAG caches.",
statePruneDur = statePruneTick - startTick,
epochRefPruneDur = epochRefPruneTick - statePruneTick
proc loadExecutionBlockRoot*(dag: ChainDAGRef, bid: BlockId): Eth2Digest =
if dag.cfg.blockForkAtEpoch(bid.slot.epoch) < BeaconBlockFork.Bellatrix:
return ZERO_HASH
let blockData = dag.getForkedBlock(bid).valueOr:
return ZERO_HASH
withBlck(blockData):
when stateFork >= BeaconStateFork.Bellatrix:
blck.message.body.execution_payload.block_hash
else:
ZERO_HASH
proc loadExecutionBlockRoot*(dag: ChainDAGRef, blck: BlockRef): Eth2Digest =
if blck.executionBlockRoot.isNone:
blck.executionBlockRoot = some dag.loadExecutionBlockRoot(blck.bid)
blck.executionBlockRoot.unsafeGet
proc updateHead*(
dag: ChainDAGRef,
newHead: BlockRef,
quarantine: var Quarantine) =
## Update what we consider to be the current head, as given by the fork
## choice.
##
## The choice of head affects the choice of finalization point - the order
## of operations naturally becomes important here - after updating the head,
## blocks that were once considered potential candidates for a tree will
## now fall from grace, or no longer be considered resolved.
doAssert not newHead.isNil()
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
# Could happen if enough blocks get invalidated and would corrupt database -
# When finalized checkpoint is empty, the slot may also be smaller
doAssert newHead.slot >= dag.finalizedHead.slot or
newHead == dag.finalizedHead.blck
let
lastHead = dag.head
logScope:
newHead = shortLog(newHead)
lastHead = shortLog(lastHead)
if lastHead == newHead:
trace "No head block update"
return
if newHead.parent.isNil:
# The new head should always have the finalizedHead as ancestor - thus,
# this should not happen except in a race condition where the selected
# `BlockRef` had its parent set to nil as happens during finalization -
# notably, resetting the head to be the finalizedHead is not allowed
error "Cannot update head to block without parent"
return
let
lastHeadStateRoot = getStateRoot(dag.headState)
lastHeadMergeComplete = dag.headState.is_merge_transition_complete()
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
# Start off by making sure we have the right state - updateState will try
# to use existing in-memory states to make this smooth
var cache: StateCache
if not updateState(
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag, dag.headState, newHead.bid.atSlot(), false, cache):
# Advancing the head state should never fail, given that the tail is
# implicitly finalised, the head is an ancestor of the tail and we always
# store the tail state in the database, as well as every epoch slot state in
# between
fatal "Unable to load head state during head update, database corrupt?",
lastHead = shortLog(lastHead)
quit 1
dag.head = newHead
if dag.headState.is_merge_transition_complete() and not
lastHeadMergeComplete and
dag.vanityLogs.onMergeTransitionBlock != nil:
dag.vanityLogs.onMergeTransitionBlock()
dag.db.putHeadBlock(newHead.root)
updateBeaconMetrics(dag.headState, dag.head.bid, cache)
withState(dag.headState):
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
when stateFork >= BeaconStateFork.Altair:
dag.headSyncCommittees = forkyState.data.get_sync_committee_cache(cache)
Speed up altair block processing 2x (#3115) * Speed up altair block processing >2x Like #3089, this PR drastially speeds up historical REST queries and other long state replays. * cache sync committee validator indices * use ~80mb less memory for validator pubkey mappings * batch-verify sync aggregate signature (fixes #2985) * document sync committee hack with head block vs sync message block * add batch signature verification failure tests Before: ``` ../env.sh nim c -d:release -r ncli_db --db:mainnet_0/db bench --start-slot:-1000 All time are ms Average, StdDev, Min, Max, Samples, Test Validation is turned off meaning that no BLS operations are performed 5830.675, 0.000, 5830.675, 5830.675, 1, Initialize DB 0.481, 1.878, 0.215, 59.167, 981, Load block from database 8422.566, 0.000, 8422.566, 8422.566, 1, Load state from database 6.996, 1.678, 0.042, 14.385, 969, Advance slot, non-epoch 93.217, 8.318, 84.192, 122.209, 32, Advance slot, epoch 20.513, 23.665, 11.510, 201.561, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` After: ``` 7081.422, 0.000, 7081.422, 7081.422, 1, Initialize DB 0.553, 2.122, 0.175, 66.692, 981, Load block from database 5439.446, 0.000, 5439.446, 5439.446, 1, Load state from database 6.829, 1.575, 0.043, 12.156, 969, Advance slot, non-epoch 94.716, 2.749, 88.395, 100.026, 32, Advance slot, epoch 11.636, 23.766, 4.889, 205.250, 981, Apply block, no slot processing 0.000, 0.000, 0.000, 0.000, 0, Database load 0.000, 0.000, 0.000, 0.000, 0, Database store ``` * add comment
2021-11-24 12:43:50 +00:00
let
finalized_checkpoint =
getStateField(dag.headState, finalized_checkpoint)
finalizedSlot =
# finalized checkpoint may move back in the head state compared to what
# we've seen in other forks - it does not move back in fork choice
# however, so we'll use the last-known-finalized in that case
max(finalized_checkpoint.epoch.start_slot(), dag.finalizedHead.slot)
finalizedHead = newHead.atSlot(finalizedSlot)
doAssert (not finalizedHead.blck.isNil),
"Block graph should always lead to a finalized block"
# Update light client data
dag.processHeadChangeForLightClient()
let (isAncestor, ancestorDepth) = lastHead.getDepth(newHead)
if not(isAncestor):
notice "Updated head block with chain reorg",
headParent = shortLog(newHead.parent),
stateRoot = shortLog(getStateRoot(dag.headState)),
justified = shortLog(getStateField(
dag.headState, current_justified_checkpoint)),
finalized = shortLog(getStateField(dag.headState, finalized_checkpoint)),
isOptHead = dag.is_optimistic(newHead.root)
if not(isNil(dag.onReorgHappened)):
let
# TODO (cheatfate): Proper implementation required
data = ReorgInfoObject.init(dag.head.slot, uint64(ancestorDepth),
lastHead.root, newHead.root,
lastHeadStateRoot,
getStateRoot(dag.headState))
dag.onReorgHappened(data)
# A reasonable criterion for "reorganizations of the chain"
quarantine.clearAfterReorg()
beacon_reorgs_total_total.inc()
beacon_reorgs_total.inc()
else:
debug "Updated head block",
stateRoot = shortLog(getStateRoot(dag.headState)),
justified = shortLog(getStateField(
dag.headState, current_justified_checkpoint)),
finalized = shortLog(getStateField(dag.headState, finalized_checkpoint)),
isOptHead = dag.is_optimistic(newHead.root)
if not(isNil(dag.onHeadChanged)):
let
currentEpoch = epoch(newHead.slot)
depRoot = withState(dag.headState): forkyState.proposer_dependent_root
prevDepRoot = withState(dag.headState):
forkyState.attester_dependent_root
epochTransition = (finalizedHead != dag.finalizedHead)
# TODO (cheatfate): Proper implementation required
data = HeadChangeInfoObject.init(dag.head.slot, dag.head.root,
getStateRoot(dag.headState),
epochTransition, prevDepRoot,
depRoot)
dag.onHeadChanged(data)
withState(dag.headState):
# Every time the head changes, the "canonical" view of balances and other
# state-related metrics change - notify the validator monitor.
# Doing this update during head update ensures there's a reasonable number
# of such updates happening - at most once per valid block.
dag.validatorMonitor[].registerState(forkyState.data)
if finalizedHead != dag.finalizedHead:
debug "Reached new finalization checkpoint",
stateRoot = shortLog(getStateRoot(dag.headState)),
Logging and startup improvements (#3038) * Logging and startup improvements Color support for released binaries! * startup scripts no longer log to file by default - this only affects source builds - released binaries don't support file logging * add --log-stdout option to control logging to stdout (colors, json) * detect tty:s vs redirected logs and log accordingly * add option to disable log colors at runtime * simplify several "common" logs, showing the most important information earlier and more clearly * remove line numbers / file information / tid - these take up space and are of little use to end users * still enabled in debug builds and tools * remove `testnet_servers_image` compile-time option * server images, released binaries and compile-from-source now offer the same behaviour and features * fixes https://github.com/status-im/nimbus-eth2/issues/2326 * fixes https://github.com/status-im/nimbus-eth2/issues/1794 * remove instanteneous block speed from sync message, keeping only average before: ``` INF 2021-10-28 16:45:59.000+02:00 Slot start topics="beacnde" tid=386429 file=nimbus_beacon_node.nim:884 lastSlot=2384027 wallSlot=2384028 delay=461us84ns peers=0 head=75a10ee5:3348 headEpoch=104 finalized=cd6804ba:3264 finalizedEpoch=102 sync="wwwwwwwwww:0:0.0000:0.0000:00h00m (3348)" INF 2021-10-28 16:45:59.046+02:00 Slot end topics="beacnde" tid=386429 file=nimbus_beacon_node.nim:821 slot=2384028 nextSlot=2384029 head=75a10ee5:3348 headEpoch=104 finalizedHead=cd6804ba:3264 finalizedEpoch=102 nextAttestationSlot=-1 nextProposalSlot=-1 nextActionWait=n/a ``` after: ``` INF 2021-10-28 22:43:23.033+02:00 Slot start topics="beacnde" slot=2385815 epoch=74556 sync="DDPDDPUDDD:10:5.2258:01h19m (2361088)" peers=37 head=eacd2dae:2361096 finalized=73782:a4751487 delay=33ms687us715ns INF 2021-10-28 22:43:23.291+02:00 Slot end topics="beacnde" slot=2385815 nextActionWait=n/a nextAttestationSlot=-1 nextProposalSlot=-1 head=eacd2dae:2361096 ``` * fix comment * documentation updates * mention `--log-file` may be deprecated in the future * update various docs
2021-11-02 17:06:36 +00:00
justified = shortLog(getStateField(
dag.headState, current_justified_checkpoint)),
finalized = shortLog(getStateField(dag.headState, finalized_checkpoint))
let oldFinalizedHead = dag.finalizedHead
block:
# Update `dag.finalizedBlocks` with all newly finalized blocks (those
# newer than the previous finalized head), then update `dag.finalizedHead`
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
var newFinalized: seq[BlockId]
var tmp = finalizedHead.blck
while not isNil(tmp) and tmp.slot >= dag.finalizedHead.slot:
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
newFinalized.add(tmp.bid)
limit by-root requests to non-finalized blocks (#3293) * limit by-root requests to non-finalized blocks Presently, we keep a mapping from block root to `BlockRef` in memory - this has simplified reasoning about the dag, but is not sustainable with the chain growing. We can distinguish between two cases where by-root access is useful: * unfinalized blocks - this is where the beacon chain is operating generally, by validating incoming data as interesting for future fork choice decisions - bounded by the length of the unfinalized period * finalized blocks - historical access in the REST API etc - no bounds, really In this PR, we limit the by-root block index to the first use case: finalized chain data can more efficiently be addressed by slot number. Future work includes: * limiting the `BlockRef` horizon in general - each instance is 40 bytes+overhead which adds up - this needs further refactoring to deal with the tail vs state problem * persisting the finalized slot-to-hash index - this one also keeps growing unbounded (albeit slowly) Anyway, this PR easily shaves ~128mb of memory usage at the time of writing. * No longer honor `BeaconBlocksByRoot` requests outside of the non-finalized period - previously, Nimbus would generously return any block through this libp2p request - per the spec, finalized blocks should be fetched via `BeaconBlocksByRange` instead. * return `Opt[BlockRef]` instead of `nil` when blocks can't be found - this becomes a lot more common now and thus deserves more attention * `dag.blocks` -> `dag.forkBlocks` - this index only carries unfinalized blocks from now - `finalizedBlocks` covers the other `BlockRef` instances * in backfill, verify that the last backfilled block leads back to genesis, or panic * add backfill timings to log * fix missing check that `BlockRef` block can be fetched with `getForkedBlock` reliably * shortcut doppelganger check when feature is not enabled * in REST/JSON-RPC, fetch blocks without involving `BlockRef` * fix dag.blocks ref
2022-01-21 11:33:16 +00:00
if tmp != finalizedHead.blck:
# The newly finalized block itself should remain in here so that fork
# choice still can find it via root
dag.forkBlocks.excl(KeyedBlockRef.init(tmp))
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
let p = tmp.parent
tmp.parent = nil # Reset all parent links to release memory
tmp = p
dag.finalizedHead = finalizedHead
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
dag.db.updateFinalizedBlocks(newFinalized)
Store finalized block roots in database (3s startup) (#3320) * Store finalized block roots in database (3s startup) When the chain has finalized a checkpoint, the history from that point onwards becomes linear - this is exploited in `.era` files to allow constant-time by-slot lookups. In the database, we can do the same by storing finalized block roots in a simple sparse table indexed by slot, bringing the two representations closer to each other in terms of conceptual layout and performance. Doing so has a number of interesting effects: * mainnet startup time is improved 3-5x (3s on my laptop) * the _first_ startup might take slightly longer as the new index is being built - ~10s on the same laptop * we no longer rely on the beacon block summaries to load the full dag - this is a lot faster because we no longer have to look up each block by parent root * a collateral benefit is that we no longer need to load the full summaries table into memory - we get the RSS benefits of #3164 without the CPU hit. Other random stuff: * simplify forky block generics * fix withManyWrites multiple evaluation * fix validator key cache not being updated properly in chaindag read-only mode * drop pre-altair summaries from `kvstore` * recreate missing summaries from altair+ blocks as well (in case database has lost some to an involuntary restart) * print database startup timings in chaindag load log * avoid allocating superfluos state at startup * use a recursive sql query to load the summaries of the unfinalized blocks
2022-01-30 16:51:04 +00:00
if dag.loadExecutionBlockRoot(oldFinalizedHead.blck).isZero and
not dag.loadExecutionBlockRoot(dag.finalizedHead.blck).isZero and
dag.vanityLogs.onFinalizedMergeTransitionBlock != nil:
dag.vanityLogs.onFinalizedMergeTransitionBlock()
# Pruning the block dag is required every time the finalized head changes
# in order to clear out blocks that are no longer viable and should
# therefore no longer be considered as part of the chain we're following
dag.pruneBlocksDAG()
# Update light client data
dag.processFinalizationForLightClient(oldFinalizedHead)
# Send notification about new finalization point via callback.
if not(isNil(dag.onFinHappened)):
let stateRoot =
if dag.finalizedHead.slot == dag.head.slot: getStateRoot(dag.headState)
elif dag.finalizedHead.slot + SLOTS_PER_HISTORICAL_ROOT > dag.head.slot:
getStateField(dag.headState, state_roots).data[
int(dag.finalizedHead.slot mod SLOTS_PER_HISTORICAL_ROOT)]
else:
Eth2Digest() # The thing that finalized was >8192 blocks old?
# TODO (cheatfate): Proper implementation required
let data = FinalizationInfoObject.init(
dag.finalizedHead.blck.root, stateRoot, dag.finalizedHead.slot.epoch)
dag.onFinHappened(dag, data)
proc isInitialized*(T: type ChainDAGRef, db: BeaconChainDB): Result[void, cstring] =
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
## Lightweight check to see if it is likely that the given database has been
## initialized
let
tailBlockRoot = db.getTailBlock()
if not tailBlockRoot.isSome():
return err("Tail block root missing")
let
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
tailBlock = db.getBlockId(tailBlockRoot.get())
if not tailBlock.isSome():
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
return err("Tail block information missing")
ok()
proc preInit*(
T: type ChainDAGRef, db: BeaconChainDB, state: ForkedHashedBeaconState) =
## Initialize a database using the given state, which potentially may be a
## non-genesis state.
##
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
## When used with a non-genesis state, the resulting database will not be
## compatible with pre-22.11 versions.
logScope:
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
stateRoot = $getStateRoot(state)
stateSlot = getStateField(state, slot)
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
doAssert getStateField(state, slot).is_epoch,
"Can only initialize database from epoch states"
withState(state):
db.putState(forkyState)
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
if forkyState.data.slot == GENESIS_SLOT:
let blck = get_initial_beacon_block(forkyState)
db.putBlock(blck)
db.putGenesisBlock(blck.root)
db.putHeadBlock(blck.root)
db.putTailBlock(blck.root)
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
notice "Database initialized from genesis",
blockRoot = $blck.root
else:
let blockRoot = forkyState.latest_block_root()
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
# We write a summary but not the block contents - these will have to be
# backfilled from the network
db.putBeaconBlockSummary(blockRoot, BeaconBlockSummary(
slot: forkyState.data.latest_block_header.slot,
parent_root: forkyState.data.latest_block_header.parent_root
))
db.putHeadBlock(blockRoot)
db.putTailBlock(blockRoot)
State-only checkpoint state startup (#4251) Currently, we require genesis and a checkpoint block and state to start from an arbitrary slot - this PR relaxes this requirement so that we can start with a state alone. The current trusted-node-sync algorithm works by first downloading blocks until we find an epoch aligned non-empty slot, then downloads the state via slot. However, current [proposals](https://github.com/ethereum/beacon-APIs/pull/226) for checkpointing prefer finalized state as the main reference - this allows more simple access control and caching on the server side - in particular, this should help checkpoint-syncing from sources that have a fast `finalized` state download (like infura and teku) but are slow when accessing state via slot. Earlier versions of Nimbus will not be able to read databases created without a checkpoint block and genesis. In most cases, backfilling makes the database compatible except where genesis is also missing (custom networks). * backfill checkpoint block from libp2p instead of checkpoint source, when doing trusted node sync * allow starting the client without genesis / checkpoint block * perform epoch start slot lookahead when loading tail state, so as to deal with the case where the epoch start slot does not have a block * replace `--blockId` with `--state-id` in TNS command line * when replaying, also look at the parent of the last-known-block (even if we don't have the parent block data, we can still replay from a "parent" state) - in particular, this clears the way for implementing state pruning * deprecate `--finalized-checkpoint-block` option (no longer needed)
2022-11-02 10:02:38 +00:00
if db.getGenesisBlock().isSome():
notice "Checkpoint written to database", blockRoot = $blockRoot
else:
notice "Database initialized from checkpoint", blockRoot = $blockRoot
proc getProposer*(
dag: ChainDAGRef, head: BlockRef, slot: Slot): Option[ValidatorIndex] =
let
epochRef = dag.getEpochRef(head.bid, slot.epoch(), false).valueOr:
notice "Cannot load EpochRef for given head", head, slot, error
return none(ValidatorIndex)
slotInEpoch = slot.since_epoch_start()
let proposer = epochRef.beacon_proposers[slotInEpoch]
if proposer.isSome():
if proposer.get().uint64 >= dag.db.immutableValidators.lenu64():
# Sanity check - it should never happen that the key cache doesn't contain
# a key for the selected proposer - that would mean that we somehow
# created validators in the state without updating the cache!
warn "Proposer key not found",
keys = dag.db.immutableValidators.lenu64(), proposer = proposer.get()
return none(ValidatorIndex)
proposer
proc getProposalState*(
dag: ChainDAGRef, head: BlockRef, slot: Slot, cache: var StateCache):
Result[ref ForkedHashedBeaconState, cstring] =
## Return a state suitable for making proposals for the given head and slot -
## in particular, the state can be discarded after use and does not have a
## state root set
# Start with the clearance state, since this one typically has been advanced
# and thus has a hot hash tree cache
let state = newClone(dag.clearanceState)
var
info = ForkedEpochInfo()
if not state[].can_advance_slots(head.root, slot):
# The last state root will be computed as part of block production, so skip
# it now
if not dag.updateState(
state[], head.atSlot(slot - 1).toBlockSlotId().expect("not nil"),
false, cache):
error "Cannot get proposal state - skipping block production, database corrupt?",
head = shortLog(head),
slot
return err("Cannot create proposal state")
else:
loadStateCache(dag, cache, head.bid, slot.epoch)
if getStateField(state[], slot) < slot:
process_slots(
dag.cfg, state[], slot, cache, info,
{skipLastStateRootCalculation}).expect("advancing 1 slot should not fail")
ok state
proc aggregateAll*(
dag: ChainDAGRef,
validator_indices: openArray[ValidatorIndex]): Result[CookedPubKey, cstring] =
if validator_indices.len == 0:
# Aggregation spec requires non-empty collection
# - https://tools.ietf.org/html/draft-irtf-cfrg-bls-signature-04
# Eth2 spec requires at least one attesting index in attestation
# - https://github.com/ethereum/consensus-specs/blob/v1.3.0-alpha.1/specs/phase0/beacon-chain.md#is_valid_indexed_attestation
return err("aggregate: no attesting keys")
let
firstKey = dag.validatorKey(validator_indices[0])
if not firstKey.isSome():
return err("aggregate: invalid validator index")
var aggregateKey{.noinit.}: AggregatePublicKey
aggregateKey.init(firstKey.get())
for i in 1 ..< validator_indices.len:
let key = dag.validatorKey(validator_indices[i])
if not key.isSome():
return err("aggregate: invalid validator index")
aggregateKey.aggregate(key.get())
ok(finish(aggregateKey))
proc aggregateAll*(
dag: ChainDAGRef,
validator_indices: openArray[ValidatorIndex|uint64],
bits: BitSeq | BitArray): Result[CookedPubKey, cstring] =
if validator_indices.len() != bits.len():
return err("aggregateAll: mismatch in bits length")
var
aggregateKey{.noinit.}: AggregatePublicKey
inited = false
for i in 0..<bits.len():
if bits[i]:
let key = dag.validatorKey(validator_indices[i])
if not key.isSome():
return err("aggregate: invalid validator index")
if inited:
aggregateKey.aggregate(key.get)
else:
aggregateKey = AggregatePublicKey.init(key.get)
inited = true
if not inited:
err("aggregate: no attesting keys")
else:
ok(finish(aggregateKey))
func needsBackfill*(dag: ChainDAGRef): bool =
dag.backfill.slot > GENESIS_SLOT
proc rebuildIndex*(dag: ChainDAGRef) =
## After a checkpoint sync, we lack intermediate states to replay from - this
## function rebuilds them so that historical replay can take place again
## TODO handle databases without genesis state
if dag.backfill.slot > 0:
debug "Backfill not complete, cannot rebuild archive"
return
if dag.tail.slot == GENESIS_SLOT:
# The tail is the earliest slot for which we're supposed to have states -
# if it's sufficiently recent, don't do anything
debug "Archive does not need rebuilding"
return
# First, we check what states we already have in the database - that allows
# resuming the operation at any time
let
roots = dag.db.loadStateRoots()
var
canonical = newSeq[Eth2Digest](
(dag.finalizedHead.slot.epoch + EPOCHS_PER_STATE_SNAPSHOT - 1) div
EPOCHS_PER_STATE_SNAPSHOT)
# `junk` puts in place some infrastructure to prune unnecessary states - it
# will be more useful in the future as a base for pruning
junk: seq[((Slot, Eth2Digest), Eth2Digest)]
for k, v in roots:
if k[0] >= dag.finalizedHead.slot:
continue # skip newer stuff
if not isFinalizedStateSnapshot(k[0]):
# `tail` will move at the end of the process, so we won't need any
# intermediate states
junk.add((k, v))
continue # skip non-snapshot slots
if k[0] > 0:
let bs = dag.getBlockIdAtSlot(k[0] - 1)
if bs.isNone or bs.get().bid.root != k[1]:
# remove things that are no longer a canonical part of the chain or
# cannot be reached via a block
junk.add((k, v))
continue
if not dag.db.containsState(v):
continue # If it's not in the database..
canonical[k[0].epoch div EPOCHS_PER_STATE_SNAPSHOT] = v
let
state = (ref ForkedHashedBeaconState)()
var
cache: StateCache
info: ForkedEpochInfo
# `canonical` holds all slots at which a state is expected to appear, using a
# zero root whenever a particular state is missing - this way, if there's
# partial progress or gaps, they will be dealt with correctly
for i, state_root in canonical.mpairs():
if not state_root.isZero:
continue
doAssert i > 0, "Genesis should always be available"
let
startSlot = Epoch((i - 1) * EPOCHS_PER_STATE_SNAPSHOT).start_slot
slot = Epoch(i * EPOCHS_PER_STATE_SNAPSHOT).start_slot
info "Recreating state snapshot",
slot, startStateRoot = canonical[i - 1], startSlot
if getStateRoot(state[]) != canonical[i - 1]:
Prune `BlockRef` on finalization (#3513) Up til now, the block dag has been using `BlockRef`, a structure adapted for a full DAG, to represent all of chain history. This is a correct and simple design, but does not exploit the linearity of the chain once parts of it finalize. By pruning the in-memory `BlockRef` structure at finalization, we save, at the time of writing, a cool ~250mb (or 25%:ish) chunk of memory landing us at a steady state of ~750mb normal memory usage for a validating node. Above all though, we prevent memory usage from growing proportionally with the length of the chain, something that would not be sustainable over time - instead, the steady state memory usage is roughly determined by the validator set size which grows much more slowly. With these changes, the core should remain sustainable memory-wise post-merge all the way to withdrawals (when the validator set is expected to grow). In-memory indices are still used for the "hot" unfinalized portion of the chain - this ensure that consensus performance remains unchanged. What changes is that for historical access, we use a db-based linear slot index which is cache-and-disk-friendly, keeping the cost for accessing historical data at a similar level as before, achieving the savings at no percievable cost to functionality or performance. A nice collateral benefit is the almost-instant startup since we no longer load any large indicies at dag init. The cost of this functionality instead can be found in the complexity of having to deal with two ways of traversing the chain - by `BlockRef` and by slot. * use `BlockId` instead of `BlockRef` where finalized / historical data may be required * simplify clearance pre-advancement * remove dag.finalizedBlocks (~50:ish mb) * remove `getBlockAtSlot` - use `getBlockIdAtSlot` instead * `parent` and `atSlot` for `BlockId` now require a `ChainDAGRef` instance, unlike `BlockRef` traversal * prune `BlockRef` parents on finality (~200:ish mb) * speed up ChainDAG init by not loading finalized history index * mess up light client server error handling - this need revisiting :)
2022-03-17 17:42:56 +00:00
if not dag.db.getState(dag.cfg.stateForkAtEpoch(startSlot.epoch), canonical[i - 1], state[], noRollback):
error "Can't load start state, database corrupt?",
startStateRoot = shortLog(canonical[i - 1]), slot = startSlot
return
for slot in startSlot..<startSlot + (EPOCHS_PER_STATE_SNAPSHOT * SLOTS_PER_EPOCH):
let bids = dag.getBlockIdAtSlot(slot).valueOr:
warn "Block id missing, cannot continue - database corrupt?", slot
return
# The slot check is needed to avoid re-applying a block
if bids.isProposed and getStateField(state[], latest_block_header).slot < bids.bid.slot:
let res = dag.applyBlock(state[], bids.bid, cache, info)
if res.isErr:
error "Failed to apply block while building index",
state_bid = shortLog(state[].latest_block_id()),
error = res.error()
return
if slot.is_epoch:
cache.prune(slot.epoch)
process_slots(
dag.cfg, state[], slot, cache, info,
dag.updateFlags).expect("process_slots shouldn't fail when state slot is correct")
withState(state[]):
dag.db.putState(forkyState)
dag.db.checkpoint()
state_root = forkyState.root
# Now that we have states all the way to genesis, we can adjust the tail
# and readjust the in-memory indices to what they would look like if we had
# started with an earlier tail
let
genesis =
dag.getBlockIdAtSlot(GENESIS_SLOT).expect("Genesis in database").bid
dag.db.putTailBlock(genesis.root)
dag.tail = genesis
if junk.len > 0:
info "Dropping redundant states", states = junk.len
for i in junk:
dag.db.delState(i[1])