nimbus-eth2/beacon_chain/consensus_object_pools/blockchain_dag_light_client...

909 lines
37 KiB
Nim
Raw Normal View History

# beacon_chain
# Copyright (c) 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.
{.push raises: [Defect].}
# This implements the pre-release proposal of the libp2p based light client sync
# protocol. See https://github.com/ethereum/consensus-specs/pull/2802
import
# Standard library
std/[algorithm, sequtils],
# Status libraries
stew/[bitops2, objects],
chronos,
# Beacon chain internals
../spec/datatypes/[phase0, altair, bellatrix],
"."/[block_pools_types, blockchain_dag]
logScope: topics = "chaindag"
type
HashedBeaconStateWithSyncCommittee =
bellatrix.HashedBeaconState |
altair.HashedBeaconState
TrustedSignedBeaconBlockWithSyncAggregate =
bellatrix.TrustedSignedBeaconBlock |
altair.TrustedSignedBeaconBlock
template nextEpochBoundarySlot(slot: Slot): Slot =
## Compute the first possible epoch boundary state slot of a `Checkpoint`
## referring to a block at given slot.
(slot + (SLOTS_PER_EPOCH - 1)).epoch.start_slot
func computeEarliestLightClientSlot*(dag: ChainDAGRef): Slot =
## Compute the earliest slot for which light client data is retained.
let
minSupportedSlot = max(
dag.cfg.ALTAIR_FORK_EPOCH.start_slot,
dag.lightClientCache.importTailSlot)
currentSlot = getStateField(dag.headState, slot)
if currentSlot < minSupportedSlot:
return minSupportedSlot
let
MIN_EPOCHS_FOR_BLOCK_REQUESTS =
dag.cfg.MIN_VALIDATOR_WITHDRAWABILITY_DELAY +
dag.cfg.CHURN_LIMIT_QUOTIENT div 2
MIN_SLOTS_FOR_BLOCK_REQUESTS =
MIN_EPOCHS_FOR_BLOCK_REQUESTS * SLOTS_PER_EPOCH
minSlot = max(minSupportedSlot, dag.tail.slot)
if currentSlot - minSlot < MIN_SLOTS_FOR_BLOCK_REQUESTS:
return minSlot
let earliestSlot = currentSlot - MIN_SLOTS_FOR_BLOCK_REQUESTS
max(earliestSlot.sync_committee_period.start_slot, minSlot)
proc currentSyncCommitteeForPeriod(
dag: ChainDAGRef,
tmpState: var ForkedHashedBeaconState,
period: SyncCommitteePeriod): SyncCommittee =
## Fetch a `SyncCommittee` for a given sync committee period.
## For non-finalized periods, follow the chain as selected by fork choice.
let earliestSlot = dag.computeEarliestLightClientSlot
doAssert period >= earliestSlot.sync_committee_period
let
periodStartSlot = period.start_slot
syncCommitteeSlot = max(periodStartSlot, earliestSlot)
# TODO introduce error handling in the case that we don't have historical
# data for the period
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 = dag.getBlockIdAtSlot(syncCommitteeSlot).expect("TODO")
dag.withUpdatedState(tmpState, bsi) do:
withState(state):
when stateFork >= BeaconStateFork.Altair:
state.data.current_sync_committee
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
template syncCommitteeRoot(
state: HashedBeaconStateWithSyncCommittee): Eth2Digest =
## Compute a root to uniquely identify `current_sync_committee` and
## `next_sync_committee`.
withEth2Hash:
h.update state.data.current_sync_committee.hash_tree_root().data
h.update state.data.next_sync_committee.hash_tree_root().data
proc syncCommitteeRootForPeriod(
dag: ChainDAGRef,
tmpState: var ForkedHashedBeaconState,
period: SyncCommitteePeriod): Eth2Digest =
## Compute a root to uniquely identify `current_sync_committee` and
## `next_sync_committee` for a given sync committee period.
## For non-finalized periods, follow the chain as selected by fork choice.
let earliestSlot = dag.computeEarliestLightClientSlot
doAssert period >= earliestSlot.sync_committee_period
let
periodStartSlot = period.start_slot
syncCommitteeSlot = max(periodStartSlot, earliestSlot)
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 = dag.getBlockIdAtSlot(syncCommitteeSlot).expect("TODO")
dag.withUpdatedState(tmpState, bsi) do:
withState(state):
when stateFork >= BeaconStateFork.Altair:
state.syncCommitteeRoot
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
proc getLightClientData(
dag: ChainDAGRef,
bid: BlockId): CachedLightClientData =
## Fetch cached light client data about a given block.
## Data must be cached (`cacheLightClientData`) before calling this function.
try: dag.lightClientCache.data[bid]
except KeyError: raiseAssert "Unreachable"
template getLightClientData(
dag: ChainDAGRef,
blck: BlockRef): CachedLightClientData =
getLightClientData(dag, blck.bid)
proc cacheLightClientData*(
dag: ChainDAGRef,
state: HashedBeaconStateWithSyncCommittee,
blck: TrustedSignedBeaconBlockWithSyncAggregate,
isNew = true) =
## Cache data for a given block and its post-state to speed up creating future
## `LightClientUpdate` and `LightClientBootstrap` instances that refer to this
## block and state.
var current_sync_committee_branch {.noinit.}:
array[log2trunc(altair.CURRENT_SYNC_COMMITTEE_INDEX), Eth2Digest]
state.data.build_proof(
altair.CURRENT_SYNC_COMMITTEE_INDEX, current_sync_committee_branch)
var next_sync_committee_branch {.noinit.}:
array[log2trunc(altair.NEXT_SYNC_COMMITTEE_INDEX), Eth2Digest]
state.data.build_proof(
altair.NEXT_SYNC_COMMITTEE_INDEX, next_sync_committee_branch)
var finality_branch {.noinit.}:
array[log2trunc(altair.FINALIZED_ROOT_INDEX), Eth2Digest]
state.data.build_proof(
altair.FINALIZED_ROOT_INDEX, finality_branch)
template finalized_checkpoint(): auto = state.data.finalized_checkpoint
let
bid =
BlockId(root: blck.root, slot: blck.message.slot)
finalized_bid =
dag.getBlockIdAtSlot(finalized_checkpoint.epoch.start_slot).expect("TODO").bid
if dag.lightClientCache.data.hasKeyOrPut(
bid,
CachedLightClientData(
current_sync_committee_branch:
current_sync_committee_branch,
next_sync_committee_branch:
next_sync_committee_branch,
finalized_bid:
finalized_bid,
finality_branch:
finality_branch)):
doAssert false, "Redundant `cacheLightClientData` call"
proc deleteLightClientData*(dag: ChainDAGRef, bid: BlockId) =
## Delete cached light client data for a given block. This needs to be called
## when a block becomes unreachable due to finalization of a different fork.
if dag.importLightClientData == ImportLightClientData.None:
return
dag.lightClientCache.data.del bid
template lazy_header(name: untyped): untyped {.dirty.} =
## `createLightClientUpdates` helper to lazily load a known block header.
var `name ptr`: ptr[BeaconBlockHeader]
template `assign name`(target: var BeaconBlockHeader,
blockParam: BlockId | BlockRef): untyped =
if `name ptr` != nil:
target = `name ptr`[]
else:
let bid =
when blockParam is BlockId:
blockParam
else:
blockParam.bid
target = toBeaconBlockHeader(dag.getForkedBlock(bid).get)
`name ptr` = addr target
template lazy_data(name: untyped): untyped {.dirty.} =
## `createLightClientUpdates` helper to lazily load cached light client state.
var `name` {.noinit.}: CachedLightClientData
`name`.finalized_bid.slot = FAR_FUTURE_SLOT
template `load name`(bid: BlockId | BlockRef): untyped =
if `name`.finalized_bid.slot == FAR_FUTURE_SLOT:
`name` = dag.getLightClientData(bid)
proc createLightClientUpdates(
dag: ChainDAGRef,
state: HashedBeaconStateWithSyncCommittee,
blck: TrustedSignedBeaconBlockWithSyncAggregate,
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
parent: BlockId) =
## Create `LightClientUpdate` and `OptimisticLightClientUpdate` instances for
## a given block and its post-state, and keep track of best / latest ones.
## Data about the parent block's post-state and its `finalized_checkpoint`'s
## block's post-state needs to be cached (`cacheLightClientData`) before
## calling this function.
# Verify sync committee has sufficient participants
template sync_aggregate(): auto = blck.message.body.sync_aggregate
template sync_committee_bits(): auto = sync_aggregate.sync_committee_bits
let num_active_participants = countOnes(sync_committee_bits).uint64
if num_active_participants < MIN_SYNC_COMMITTEE_PARTICIPANTS:
return
# Verify attested block (parent) is recent enough and that state is available
let
earliest_slot = dag.computeEarliestLightClientSlot
attested_slot = parent.slot
if attested_slot < earliest_slot:
return
# Verify signature does not skip a sync committee period
let
signature_slot = blck.message.slot
signature_period = signature_slot.sync_committee_period
attested_period = attested_slot.sync_committee_period
if signature_period > attested_period + 1:
return
# Update to new `OptimisticLightClientUpdate` if it attests to a later slot
lazy_header(attested_header)
template optimistic_update(): auto = dag.lightClientCache.optimisticUpdate
if attested_slot > optimistic_update.attested_header.slot:
optimistic_update.attested_header
.assign_attested_header(parent)
optimistic_update.sync_aggregate =
isomorphicCast[SyncAggregate](sync_aggregate)
optimistic_update.fork_version =
state.data.fork.current_version
optimistic_update.is_signed_by_next_sync_committee =
signature_period == attested_period + 1
if dag.onOptimisticLightClientUpdate != nil:
dag.onOptimisticLightClientUpdate(optimistic_update)
# Update to new latest `LightClientUpdate` if it attests to a later slot
lazy_data(attested_data)
lazy_data(finalized_data)
lazy_header(finalized_header)
template latest_update(): auto = dag.lightClientCache.latestUpdate
if attested_slot > latest_update.attested_header.slot:
latest_update.attested_header
.assign_attested_header(parent)
latest_update.sync_aggregate =
isomorphicCast[SyncAggregate](sync_aggregate)
latest_update.fork_version =
state.data.fork.current_version
load_attested_data(parent)
let finalized_slot = attested_data.finalized_bid.slot
if finalized_slot + UPDATE_TIMEOUT < attested_slot or
finalized_slot < earliest_slot:
latest_update.finalized_header = BeaconBlockHeader()
latest_update.finality_branch.fill(Eth2Digest())
if signature_period == attested_period + 1:
latest_update.next_sync_committee = SyncCommittee()
latest_update.next_sync_committee_branch.fill(Eth2Digest())
else:
latest_update.next_sync_committee =
state.data.next_sync_committee
latest_update.next_sync_committee_branch =
attested_data.next_sync_committee_branch
else:
latest_update.finalized_header
.assign_finalized_header(attested_data.finalized_bid)
latest_update.finality_branch =
attested_data.finality_branch
if signature_period == finalized_slot.sync_committee_period + 1:
latest_update.next_sync_committee = SyncCommittee()
latest_update.next_sync_committee_branch.fill(Eth2Digest())
else:
load_finalized_data(attested_data.finalized_bid)
latest_update.next_sync_committee =
state.data.next_sync_committee
latest_update.next_sync_committee_branch =
finalized_data.next_sync_committee_branch
# Update best `LightClientUpdate` for current period if it improved
if signature_period == attested_period:
let isNextSyncCommitteeFinalized =
signature_period.start_slot <= dag.finalizedHead.slot
var best_update =
if isNextSyncCommitteeFinalized:
dag.lightClientCache.bestUpdates.getOrDefault(signature_period)
else:
let key = (signature_period, state.syncCommitteeRoot)
dag.lightClientCache.pendingBestUpdates.getOrDefault(key)
type Verdict = enum
unknown
new_update_is_worse
new_update_is_better
var verdict = unknown
# If no best update has been recorded, new update is better
template best_sync_aggregate(): auto = best_update.sync_aggregate
let best_num_active_participants =
countOnes(best_sync_aggregate.sync_committee_bits).uint64
if best_num_active_participants < MIN_SYNC_COMMITTEE_PARTICIPANTS:
verdict = new_update_is_better
else:
# If finality changes, the finalized update is better
template finalized_period_at_signature(): auto =
state.data.finalized_checkpoint.epoch.sync_committee_period
template best_attested_slot(): auto =
best_update.attested_header.slot
if best_update.finality_branch.isZeroMemory:
if (attested_slot > dag.finalizedHead.slot or
attested_slot > best_attested_slot) and
signature_period == finalized_period_at_signature:
load_attested_data(parent)
let finalized_slot = attested_data.finalized_bid.slot
if signature_period == finalized_slot.sync_committee_period and
finalized_slot >= earliest_slot:
verdict = new_update_is_better
elif attested_slot > dag.finalizedHead.slot or
attested_slot < best_attested_slot:
if signature_period == finalized_period_at_signature:
load_attested_data(parent)
let finalized_slot = attested_data.finalized_bid.slot
if signature_period != finalized_slot.sync_committee_period or
finalized_slot < earliest_slot:
verdict = new_update_is_worse
else:
verdict = new_update_is_worse
if verdict == unknown:
# If participation changes, higher participation is better
if num_active_participants < best_num_active_participants:
verdict = new_update_is_worse
elif num_active_participants > best_num_active_participants:
verdict = new_update_is_better
else:
# Older updates are better
if attested_slot >= best_attested_slot:
verdict = new_update_is_worse
else:
verdict = new_update_is_better
if verdict == new_update_is_better:
best_update.attested_header
.assign_attested_header(parent)
best_update.sync_aggregate =
isomorphicCast[SyncAggregate](sync_aggregate)
best_update.fork_version =
state.data.fork.current_version
load_attested_data(parent)
let finalized_slot = attested_data.finalized_bid.slot
if signature_period != finalized_slot.sync_committee_period or
finalized_slot < earliest_slot:
best_update.finalized_header = BeaconBlockHeader()
best_update.finality_branch.fill(Eth2Digest())
best_update.next_sync_committee =
state.data.next_sync_committee
best_update.next_sync_committee_branch =
attested_data.next_sync_committee_branch
else:
best_update.finalized_header
.assign_finalized_header(attested_data.finalized_bid)
best_update.finality_branch =
attested_data.finality_branch
load_finalized_data(attested_data.finalized_bid)
best_update.next_sync_committee =
state.data.next_sync_committee
best_update.next_sync_committee_branch =
finalized_data.next_sync_committee_branch
if isNextSyncCommitteeFinalized:
dag.lightClientCache.bestUpdates[signature_period] = best_update
debug "Best update for period improved",
period = signature_period, update = best_update
else:
let key = (signature_period, state.syncCommitteeRoot)
dag.lightClientCache.pendingBestUpdates[key] = best_update
debug "Best update for period improved",
period = key, update = best_update
proc processNewBlockForLightClient*(
dag: ChainDAGRef,
state: ForkedHashedBeaconState,
signedBlock: ForkyTrustedSignedBeaconBlock,
parent: BlockRef) =
## Update light client data with information from a new block.
if dag.importLightClientData == ImportLightClientData.None:
return
if signedBlock.message.slot < dag.computeEarliestLightClientSlot:
return
when signedBlock is bellatrix.TrustedSignedBeaconBlock:
dag.cacheLightClientData(state.bellatrixData, signedBlock)
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.createLightClientUpdates(state.bellatrixData, signedBlock, parent.bid)
elif signedBlock is altair.TrustedSignedBeaconBlock:
dag.cacheLightClientData(state.altairData, signedBlock)
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.createLightClientUpdates(state.altairData, signedBlock, parent.bid)
elif signedBlock is phase0.TrustedSignedBeaconBlock:
discard
else:
{.error: "Unreachable".}
proc processHeadChangeForLightClient*(dag: ChainDAGRef) =
## Update light client data to account for a new head block.
## Note that `dag.finalizedHead` is not yet updated when this is called.
if dag.importLightClientData == ImportLightClientData.None:
return
if dag.head.slot < dag.computeEarliestLightClientSlot:
return
let headPeriod = dag.head.slot.sync_committee_period
if headPeriod.start_slot > dag.finalizedHead.slot:
let finalizedPeriod = dag.finalizedHead.slot.sync_committee_period
if headPeriod > finalizedPeriod + 1:
var tmpState = assignClone(dag.headState)
for period in finalizedPeriod + 1 ..< headPeriod:
let key = (period, dag.syncCommitteeRootForPeriod(tmpState[], period))
dag.lightClientCache.bestUpdates[period] =
dag.lightClientCache.pendingBestUpdates.getOrDefault(key)
withState(dag.headState):
when stateFork >= BeaconStateFork.Altair:
let key = (headPeriod, state.syncCommitteeRoot)
dag.lightClientCache.bestUpdates[headPeriod] =
dag.lightClientCache.pendingBestUpdates.getOrDefault(key)
else: raiseAssert "Unreachable"
proc processFinalizationForLightClient*(dag: ChainDAGRef) =
## Prune cached data that is no longer useful for creating future
## `LightClientUpdate` and `LightClientBootstrap` instances.
## This needs to be called whenever `finalized_checkpoint` changes.
if dag.importLightClientData == ImportLightClientData.None:
return
let
altairStartSlot = dag.cfg.ALTAIR_FORK_EPOCH.start_slot
finalizedSlot = dag.finalizedHead.slot
if finalizedSlot < altairStartSlot:
return
let earliestSlot = dag.computeEarliestLightClientSlot
# Keep track of latest four finalized checkpoints
let
lastIndex = dag.lightClientCache.lastCheckpointIndex
lastCheckpoint = addr dag.lightClientCache.latestCheckpoints[lastIndex]
if dag.finalizedHead.slot.epoch != lastCheckpoint.epoch or
dag.finalizedHead.blck.root != lastCheckpoint.root:
let
nextIndex = (lastIndex + 1) mod dag.lightClientCache.latestCheckpoints.len
nextCheckpoint = addr dag.lightClientCache.latestCheckpoints[nextIndex]
nextCheckpoint[].epoch = dag.finalizedHead.slot.epoch
nextCheckpoint[].root = dag.finalizedHead.blck.root
dag.lightClientCache.lastCheckpointIndex = nextIndex
# Cache `LightClientBootstrap` for newly finalized epoch boundary blocks.
# Epoch boundary blocks are the block for the initial slot of an epoch,
# or the most recent block if no block was proposed at that slot
let lowSlot = max(lastCheckpoint.epoch.start_slot, earliestSlot)
var boundarySlot = dag.finalizedHead.slot
while boundarySlot >= lowSlot:
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 bid = dag.getBlockIdAtSlot(boundarySlot).expect("historical data").bid
if bid.slot >= lowSlot:
dag.lightClientCache.bootstrap[bid.slot] =
CachedLightClientBootstrap(
current_sync_committee_branch:
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.getLightClientData(bid).current_sync_committee_branch)
boundarySlot = bid.slot.nextEpochBoundarySlot
if boundarySlot < SLOTS_PER_EPOCH:
break
boundarySlot -= SLOTS_PER_EPOCH
# Prune light client data that is no longer relevant,
# i.e., can no longer be referred to by future updates, or is too old
var bidsToDelete: seq[BlockId]
for bid, data in dag.lightClientCache.data:
if bid.slot >= earliestSlot:
if bid.slot >= finalizedSlot:
continue
if dag.lightClientCache.latestCheckpoints.anyIt(bid.root == it.root):
continue
bidsToDelete.add bid
for bid in bidsToDelete:
dag.lightClientCache.data.del bid
# Prune bootstrap data that is no longer relevant
var slotsToDelete: seq[Slot]
for slot in dag.lightClientCache.bootstrap.keys:
if slot < earliestSlot:
slotsToDelete.add slot
for slot in slotsToDelete:
dag.lightClientCache.bootstrap.del slot
# Prune best `LightClientUpdate` that are no longer relevant
let earliestPeriod = earliestSlot.sync_committee_period
var periodsToDelete: seq[SyncCommitteePeriod]
for period in dag.lightClientCache.bestUpdates.keys:
if period < earliestPeriod:
periodsToDelete.add period
for period in periodsToDelete:
dag.lightClientCache.bestUpdates.del period
# Prune best `LightClientUpdate` referring to non-finalized sync committees
# that are no longer relevant, i.e., orphaned or too old
let finalizedPeriod = finalizedSlot.sync_committee_period
var keysToDelete: seq[(SyncCommitteePeriod, Eth2Digest)]
for (period, committeeRoot) in dag.lightClientCache.pendingBestUpdates.keys:
if period <= finalizedPeriod:
keysToDelete.add (period, committeeRoot)
for key in keysToDelete:
dag.lightClientCache.pendingBestUpdates.del key
proc initBestLightClientUpdateForPeriod(
dag: ChainDAGRef, period: SyncCommitteePeriod) =
## Compute and cache the `LightClientUpdate` with the most sync committee
## signatures (i.e., participation) for a given sync committee period.
let periodStartSlot = period.start_slot
if periodStartSlot > dag.finalizedHead.slot:
return
let
earliestSlot = dag.computeEarliestLightClientSlot
periodEndSlot = periodStartSlot + SLOTS_PER_SYNC_COMMITTEE_PERIOD - 1
if periodEndSlot < earliestSlot:
return
if dag.lightClientCache.bestUpdates.hasKey(period):
return
let startTick = Moment.now()
debug "Computing best update for period", period
defer:
let endTick = Moment.now()
debug "Best update for period computed",
period, update = dag.lightClientCache.bestUpdates.getOrDefault(period),
computeDur = endTick - startTick
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 maxParticipantsBlock(
dag: ChainDAGRef, highBlck: BlockId, lowSlot: Slot): Opt[BlockId] =
## Determine the earliest block with most sync committee signatures among
## ancestors of `highBlck` with at least `lowSlot` as parent block slot.
## Return `nil` if no block with `MIN_SYNC_COMMITTEE_PARTICIPANTS` is found.
var
maxParticipants = 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
maxBlockRef: Opt[BlockId]
blockRef = highBlck
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
while true:
let parent = dag.parent(blockRef).valueOr:
break
if parent.slot < lowSlot:
break
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
bdata = dag.getForkedBlock(blockRef).get
numParticipants =
withBlck(bdata):
when stateFork >= BeaconStateFork.Altair:
countOnes(blck.message.body.sync_aggregate.sync_committee_bits)
else: raiseAssert "Unreachable"
if numParticipants >= maxParticipants:
maxParticipants = numParticipants
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
maxBlockRef = ok blockRef
blockRef = parent
if maxParticipants < MIN_SYNC_COMMITTEE_PARTICIPANTS:
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
maxBlockRef.reset()
maxBlockRef
# Determine the block in the period with highest sync committee participation
let
lowSlot = max(periodStartSlot, earliestSlot)
highSlot = min(periodEndSlot, dag.finalizedHead.blck.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
highBlck = dag.getBlockIdAtSlot(highSlot).expect("TODO").bid
bestNonFinalizedRef = dag.maxParticipantsBlock(highBlck, lowSlot)
if bestNonFinalizedRef.isNone:
dag.lightClientCache.bestUpdates[period] = default(altair.LightClientUpdate)
return
# The block with highest participation may refer to a `finalized_checkpoint`
# in a different sync committee period. If that is the case, search for a
# later block with a `finalized_checkpoint` within the given sync committee
# period, despite it having a lower sync committee participation
var
tmpState = assignClone(dag.headState)
bestFinalizedRef = bestNonFinalizedRef
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
finalizedBlck: Opt[BlockId]
while bestFinalizedRef.isSome:
let
finalizedEpoch = 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
dag.withUpdatedState(tmpState[], dag.parent(bestFinalizedRef.get()).expect("TODO").atSlot) do:
withState(state):
when stateFork >= BeaconStateFork.Altair:
state.data.finalized_checkpoint.epoch
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
finalizedEpochStartSlot = finalizedEpoch.start_slot
if finalizedEpochStartSlot >= lowSlot:
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
finalizedBlck = Opt[BlockId].ok(dag.getBlockIdAtSlot(finalizedEpochStartSlot).expect(
"TODO").bid)
if finalizedBlck.get.slot >= lowSlot:
break
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
bestFinalizedRef = dag.maxParticipantsBlock(highBlck, bestFinalizedRef.get().slot + 1)
# If a finalized block has been found within the sync commitee period,
# create a `LightClientUpdate` for that one. Otherwise, create a non-finalized
# `LightClientUpdate`
var update {.noinit.}: LightClientUpdate
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 bestFinalizedRef.isSome:
# Fill data from attested 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
dag.withUpdatedState(tmpState[], dag.parent(bestFinalizedRef.get()).expect("TODO").atSlot) do:
let bdata = dag.getForkedBlock(bid).get
withStateAndBlck(state, bdata):
when stateFork >= BeaconStateFork.Altair:
update.attested_header =
blck.toBeaconBlockHeader
state.data.build_proof(
altair.FINALIZED_ROOT_INDEX, update.finality_branch)
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
# Fill data from signature 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 bdata = dag.getForkedBlock(bestFinalizedRef.get()).get
withBlck(bdata):
when stateFork >= BeaconStateFork.Altair:
update.sync_aggregate =
isomorphicCast[SyncAggregate](blck.message.body.sync_aggregate)
else: raiseAssert "Unreachable"
update.fork_version =
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.cfg.forkAtEpoch(bestFinalizedRef.get().slot.epoch).current_version
# Fill data from finalized 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
dag.withUpdatedState(tmpState[], finalizedBlck.get().atSlot) do:
let bdata = dag.getForkedBlock(bid).get
withStateAndBlck(state, bdata):
when stateFork >= BeaconStateFork.Altair:
update.next_sync_committee =
state.data.next_sync_committee
state.data.build_proof(
altair.NEXT_SYNC_COMMITTEE_INDEX, update.next_sync_committee_branch)
update.finalized_header =
blck.toBeaconBlockHeader
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
else:
# Fill data from attested 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
dag.withUpdatedState(tmpState[], dag.parent(bestNonFinalizedRef.get()).expect("TODO").atSlot) do:
let bdata = dag.getForkedBlock(bid).get
withStateAndBlck(state, bdata):
when stateFork >= BeaconStateFork.Altair:
update.attested_header =
blck.toBeaconBlockHeader
update.next_sync_committee =
state.data.next_sync_committee
state.data.build_proof(
altair.NEXT_SYNC_COMMITTEE_INDEX, update.next_sync_committee_branch)
update.finalized_header = BeaconBlockHeader()
update.finality_branch.fill(Eth2Digest())
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
# Fill data from signature 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 bdata = dag.getForkedBlock(bestNonFinalizedRef.get()).get
withBlck(bdata):
when stateFork >= BeaconStateFork.Altair:
update.sync_aggregate =
isomorphicCast[SyncAggregate](blck.message.body.sync_aggregate)
else: raiseAssert "Unreachable"
update.fork_version =
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.cfg.forkAtEpoch(bestNonFinalizedRef.get.slot.epoch).current_version
dag.lightClientCache.bestUpdates[period] = update
proc initLightClientBootstrapForPeriod(
dag: ChainDAGRef,
period: SyncCommitteePeriod) =
## Compute and cache `LightClientBootstrap` data for all epoch boundary blocks
## within a given sync committee period.
let periodStartSlot = period.start_slot
if periodStartSlot > dag.finalizedHead.slot:
return
let
earliestSlot = dag.computeEarliestLightClientSlot
periodEndSlot = periodStartSlot + SLOTS_PER_SYNC_COMMITTEE_PERIOD - 1
if periodEndSlot < earliestSlot:
return
let startTick = Moment.now()
debug "Caching bootstrap data for period", period
defer:
let endTick = Moment.now()
debug "Bootstrap data for period cached", period,
cacheDur = endTick - startTick
let
lowSlot = max(periodStartSlot, earliestSlot)
highSlot = min(periodEndSlot, dag.finalizedHead.blck.slot)
lowBoundarySlot = lowSlot.nextEpochBoundarySlot
highBoundarySlot = highSlot.nextEpochBoundarySlot
var
tmpState = assignClone(dag.headState)
tmpCache: StateCache
nextBoundarySlot = lowBoundarySlot
while nextBoundarySlot <= highBoundarySlot:
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
blck = dag.getBlockIdAtSlot(nextBoundarySlot).expect("TODO").bid
boundarySlot = blck.slot.nextEpochBoundarySlot
if boundarySlot == nextBoundarySlot and
blck.slot >= lowSlot and blck.slot <= highSlot and
not dag.lightClientCache.bootstrap.hasKey(blck.slot):
var cachedBootstrap {.noinit.}: CachedLightClientBootstrap
doAssert dag.updateState(
tmpState[], blck.atSlot, save = false, tmpCache)
withStateVars(tmpState[]):
withState(state):
when stateFork >= BeaconStateFork.Altair:
state.data.build_proof(
altair.CURRENT_SYNC_COMMITTEE_INDEX,
cachedBootstrap.current_sync_committee_branch)
else: raiseAssert "Unreachable"
dag.lightClientCache.bootstrap[blck.slot] = cachedBootstrap
nextBoundarySlot += SLOTS_PER_EPOCH
proc initLightClientCache*(dag: ChainDAGRef) =
## Initialize cached light client data
if dag.importLightClientData == ImportLightClientData.None:
return
if dag.importLightClientData == ImportLightClientData.OnlyNew:
dag.lightClientCache.importTailSlot = dag.head.slot + 1
return
let earliestSlot = dag.computeEarliestLightClientSlot
if dag.head.slot < earliestSlot:
return
let
finalizedSlot = dag.finalizedHead.slot
finalizedPeriod = finalizedSlot.sync_committee_period
dag.initBestLightClientUpdateForPeriod(finalizedPeriod)
let lightClientStartTick = Moment.now()
debug "Initializing cached light client data"
# Build lists of block to process.
# As it is slow to load states in descending order,
# first build a todo list, then process them in ascending order
let lowSlot = max(finalizedSlot, dag.computeEarliestLightClientSlot)
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
blocksBetween = newSeqOfCap[BlockId](dag.head.slot - lowSlot + 1)
blockRef = dag.head.bid
while blockRef.slot > lowSlot:
blocksBetween.add blockRef
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
blockRef = dag.parent(blockRef).expect("TODO")
blocksBetween.add blockRef
# Process blocks (reuses `dag.headState`, but restores it to the current head)
var
tmpState = assignClone(dag.headState)
tmpCache, cache: StateCache
oldCheckpoint: Checkpoint
cpIndex = 0
for i in countdown(blocksBetween.high, blocksBetween.low):
blockRef = blocksBetween[i]
doAssert dag.updateState(
dag.headState, blockRef.atSlot(), save = false, cache)
withStateVars(dag.headState):
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 bdata = dag.getForkedBlock(blockRef).get
withStateAndBlck(state, bdata):
when stateFork >= BeaconStateFork.Altair:
# Cache data for `LightClientUpdate` of descendant blocks
dag.cacheLightClientData(state, blck, isNew = false)
# Cache data for the block's `finalized_checkpoint`.
# The `finalized_checkpoint` may refer to:
# 1. `finalizedHead.blck -> finalized_checkpoint`
# This may happen when there were skipped slots.
# 2. `finalizedHead -> finalized_checkpoint`
# 3. One epoch boundary that got justified then finalized
# between `finalizedHead -> finalized_checkpoint`
# and `finalizedHead`
# 4. `finalizedHead`
let checkpoint = state.data.finalized_checkpoint
if checkpoint != oldCheckpoint:
oldCheckpoint = checkpoint
doAssert cpIndex < dag.lightClientCache.latestCheckpoints.len
dag.lightClientCache.latestCheckpoints[cpIndex] = checkpoint
dag.lightClientCache.lastCheckpointIndex = cpIndex
inc cpIndex
# Save new checkpoint block using `tmpState` (avoid replay after it)
# Note that light client finality proofs refer to checkpoint blocks
# at their original slot, without advancing to an epoch boundary.
# This is because light clients are unable to advance slots.
if checkpoint.root != dag.finalizedHead.blck.root:
let cpRef =
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.getBlockIdAtSlot(checkpoint.epoch.start_slot).expect("TODO").bid
if cpRef.slot >= earliestSlot:
assert cpRef.root == checkpoint.root
doAssert dag.updateState(
tmpState[], cpRef.atSlot, save = false, tmpCache)
withStateVars(tmpState[]):
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 bdata = dag.getForkedBlock(cpRef).get
withStateAndBlck(state, bdata):
when stateFork >= BeaconStateFork.Altair:
dag.cacheLightClientData(state, blck, isNew = false)
else: raiseAssert "Unreachable"
# Create `LightClientUpdate` for non-finalized blocks.
if blockRef.slot > finalizedSlot:
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.createLightClientUpdates(state, blck, dag.parent(blockRef).expect("TODO"))
else: raiseAssert "Unreachable"
let lightClientEndTick = Moment.now()
debug "Initialized cached light client data",
initDur = lightClientEndTick - lightClientStartTick
# Import historic data
if dag.importLightClientData == ImportLightClientData.Full:
let
earliestSlot = dag.computeEarliestLightClientSlot
earliestPeriod = earliestSlot.sync_committee_period
for period in earliestPeriod ..< finalizedPeriod:
dag.initBestLightClientUpdateForPeriod(period)
dag.initLightClientBootstrapForPeriod(period)
dag.initLightClientBootstrapForPeriod(finalizedPeriod)
proc getBestLightClientUpdateForPeriod*(
dag: ChainDAGRef,
period: SyncCommitteePeriod): Option[altair.LightClientUpdate] =
if not dag.serveLightClientData:
return none(altair.LightClientUpdate)
if dag.importLightClientData == ImportLightClientData.OnDemand:
dag.initBestLightClientUpdateForPeriod(period)
result = some(dag.lightClientCache.bestUpdates.getOrDefault(period))
let numParticipants = countOnes(result.get.sync_aggregate.sync_committee_bits)
if numParticipants < MIN_SYNC_COMMITTEE_PARTICIPANTS:
result = none(altair.LightClientUpdate)
proc getLatestLightClientUpdate*(
dag: ChainDAGRef): Option[altair.LightClientUpdate] =
if not dag.serveLightClientData:
return none(altair.LightClientUpdate)
result = some(dag.lightClientCache.latestUpdate)
let numParticipants = countOnes(result.get.sync_aggregate.sync_committee_bits)
if numParticipants < MIN_SYNC_COMMITTEE_PARTICIPANTS:
result = none(altair.LightClientUpdate)
proc getOptimisticLightClientUpdate*(
dag: ChainDAGRef): Option[OptimisticLightClientUpdate] =
if not dag.serveLightClientData:
return none(OptimisticLightClientUpdate)
result = some(dag.lightClientCache.optimisticUpdate)
let numParticipants = countOnes(result.get.sync_aggregate.sync_committee_bits)
if numParticipants < MIN_SYNC_COMMITTEE_PARTICIPANTS:
result = none(OptimisticLightClientUpdate)
proc getLightClientBootstrap*(
dag: ChainDAGRef,
blockRoot: Eth2Digest): Option[altair.LightClientBootstrap] =
if not dag.serveLightClientData:
return none(altair.LightClientBootstrap)
let bdata = dag.getForkedBlock(blockRoot).valueOr:
debug "Bootstrap unavailable: Block not found", blockRoot
return none(altair.LightClientBootstrap)
withBlck(bdata):
let
slot = blck.message.slot
period = slot.sync_committee_period
when stateFork >= BeaconStateFork.Altair:
let earliestSlot = dag.computeEarliestLightClientSlot
if slot < earliestSlot:
debug "Bootstrap unavailable: Block too old", slot
return none(altair.LightClientBootstrap)
if slot > dag.finalizedHead.blck.slot:
debug "Bootstrap unavailable: Not finalized", blockRoot
return none(altair.LightClientBootstrap)
var cachedBootstrap =
dag.lightClientCache.bootstrap.getOrDefault(slot)
if cachedBootstrap.current_sync_committee_branch.isZeroMemory:
if dag.importLightClientData == ImportLightClientData.OnDemand:
var tmpState = assignClone(dag.headState)
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.withUpdatedState(tmpState[], dag.getBlockIdAtSlot(slot).expect("TODO")) do:
withState(state):
when stateFork >= BeaconStateFork.Altair:
state.data.build_proof(
altair.CURRENT_SYNC_COMMITTEE_INDEX,
cachedBootstrap.current_sync_committee_branch)
else: raiseAssert "Unreachable"
do: raiseAssert "Unreachable"
dag.lightClientCache.bootstrap[slot] = cachedBootstrap
else:
debug "Bootstrap unavailable: Data not cached", slot
return none(altair.LightClientBootstrap)
var tmpState = assignClone(dag.headState)
var bootstrap {.noinit.}: altair.LightClientBootstrap
bootstrap.header =
blck.toBeaconBlockHeader
bootstrap.current_sync_committee =
dag.currentSyncCommitteeForPeriod(tmpState[], period)
bootstrap.current_sync_committee_branch =
cachedBootstrap.current_sync_committee_branch
return some(bootstrap)
else:
debug "Bootstrap unavailable: Block before Altair", slot
return none(altair.LightClientBootstrap)