nimbus-eth1/nimbus/db/aristo/aristo_init/memory_db.nim

265 lines
7.8 KiB
Nim
Raw Normal View History

# nimbus-eth1
# Copyright (c) 2023-2024 Status Research & Development GmbH
# Licensed under either of
# * Apache License, version 2.0, ([LICENSE-APACHE](LICENSE-APACHE) or
# http://www.apache.org/licenses/LICENSE-2.0)
# * MIT license ([LICENSE-MIT](LICENSE-MIT) or
# http://opensource.org/licenses/MIT)
# at your option. This file may not be copied, modified, or distributed
# except according to those terms.
## In-memory backend for Aristo DB
## ===============================
##
## The iterators provided here are currently available only by direct
## backend access
## ::
## import
## aristo/aristo_init,
## aristo/aristo_init/aristo_memory
##
## let rc = newAristoDbRef(BackendMemory)
## if rc.isOk:
## let be = rc.value.to(MemBackendRef)
## for (n, key, vtx) in be.walkVtx:
## ...
##
{.push raises: [].}
import
std/[algorithm, options, sequtils, tables],
eth/common,
results,
../aristo_constants,
../aristo_desc,
../aristo_desc/desc_backend,
Aristo db update for short nodes key edge cases (#1887) * Aristo: Provide key-value list signature calculator detail: Simple wrappers around `Aristo` core functionality * Update new API for `CoreDb` details: + Renamed new API functions `contains()` => `hasKey()` or `hasPath()` which disables the `in` operator on non-boolean `contains()` functions + The functions `get()` and `fetch()` always return a not-found error if there is no item, available. The new functions `getOrEmpty()` and `mergeOrEmpty()` return an an empty `Blob` if there is no such key found. * Rewrite `core_apps.nim` using new API from `CoreDb` * Use `Aristo` functionality for calculating Merkle signatures details: For debugging, the `VerifyAristoForMerkleRootCalc` can be set so that `Aristo` results will be verified against the legacy versions. * Provide general interface for Merkle signing key-value tables details: Export `Aristo` wrappers * Activate `CoreDb` tests why: Now, API seems to be stable enough for general tests. * Update `toHex()` usage why: Byteutils' `toHex()` is superior to `toSeq.mapIt(it.toHex(2)).join` * Split `aristo_transcode` => `aristo_serialise` + `aristo_blobify` why: + Different modules for different purposes + `aristo_serialise`: RLP encoding/decoding + `aristo_blobify`: Aristo database encoding/decoding * Compacted representation of small nodes' links instead of Keccak hashes why: Ethereum MPTs use Keccak hashes as node links if the size of an RLP encoded node is at least 32 bytes. Otherwise, the RLP encoded node value is used as a pseudo node link (rather than a hash.) Such a node is nor stored on key-value database. Rather the RLP encoded node value is stored instead of a lode link in a parent node instead. Only for the root hash, the top level node is always referred to by the hash. This feature needed an abstraction of the `HashKey` object which is now either a hash or a blob of length at most 31 bytes. This leaves two ways of representing an empty/void `HashKey` type, either as an empty blob of zero length, or the hash of an empty blob. * Update `CoreDb` interface (mainly reducing logger noise) * Fix copyright years (to make `Lint` happy)
2023-11-08 12:18:32 +00:00
../aristo_blobify,
./init_common
const
extraTraceMessages = false # or true
## Enabled additional logging noise
type
MemDbRef = ref object
## Database
sTab: Table[RootedVertexID,seq[byte]] ## Structural vertex table making up a trie
tUvi: Option[VertexID] ## Top used vertex ID
lSst: Opt[SavedState] ## Last saved state
MemBackendRef* = ref object of TypedBackendRef
## Inheriting table so access can be extended for debugging purposes
mdb: MemDbRef ## Database
MemPutHdlRef = ref object of TypedPutHdlRef
sTab: Table[RootedVertexID,seq[byte]]
2024-06-04 15:05:13 +00:00
tUvi: Option[VertexID]
lSst: Opt[SavedState]
when extraTraceMessages:
import chronicles
logScope:
topics = "aristo-backend"
# ------------------------------------------------------------------------------
# Private helpers
# ------------------------------------------------------------------------------
proc newSession(db: MemBackendRef): MemPutHdlRef =
new result
result.TypedPutHdlRef.beginSession db
proc getSession(hdl: PutHdlRef; db: MemBackendRef): MemPutHdlRef =
hdl.TypedPutHdlRef.verifySession db
hdl.MemPutHdlRef
proc endSession(hdl: PutHdlRef; db: MemBackendRef): MemPutHdlRef =
hdl.TypedPutHdlRef.finishSession db
hdl.MemPutHdlRef
# ------------------------------------------------------------------------------
# Private functions: interface
# ------------------------------------------------------------------------------
proc getVtxFn(db: MemBackendRef): GetVtxFn =
result =
proc(rvid: RootedVertexID, flags: set[GetVtxFlag]): Result[VertexRef,AristoError] =
# Fetch serialised data record
let data = db.mdb.sTab.getOrDefault(rvid, EmptyBlob)
if 0 < data.len:
let rc = data.deblobify(VertexRef)
when extraTraceMessages:
if rc.isErr:
trace logTxt "getVtxFn() failed", error=rc.error
return rc
err(GetVtxNotFound)
proc getKeyFn(db: MemBackendRef): GetKeyFn =
result =
proc(rvid: RootedVertexID): Result[HashKey,AristoError] =
Store keys together with node data (#2849) Currently, computed hash keys are stored in a separate column family with respect to the MPT data they're generated from - this has several disadvantages: * A lot of space is wasted because the lookup key (`RootedVertexID`) is repeated in both tables - this is 30% of the `AriKey` content! * rocksdb must maintain in-memory bloom filters and LRU caches for said keys, doubling its "minimal efficient cache size" * An extra disk traversal must be made to check for existence of cached hash key * Doubles the amount of files on disk due to each column family being its own set of files Here, the two CFs are joined such that both key and data is stored in `AriVtx`. This means: * we save ~30% disk space on repeated lookup keys * we save ~2gb of memory overhead that can be used to cache data instead of indices * we can skip storing hash keys for MPT leaf nodes - these are trivial to compute and waste a lot of space - previously they had to present in the `AriKey` CF to avoid having to look in two tables on the happy path. * There is a small increase in write amplification because when a hash value is updated for a branch node, we must write both key and branch data - previously we would write only the key * There's a small shift in CPU usage - instead of performing lookups in the database, hashes for leaf nodes are (re)-computed on the fly * We can return to slightly smaller on-disk SST files since there's fewer of them, which should reduce disk traffic a bit Internally, there are also other advantages: * when clearing keys, we no longer have to store a zero hash in memory - instead, we deduce staleness of the cached key from the presence of an updated VertexRef - this saves ~1gb of mem overhead during import * hash key cache becomes dedicated to branch keys since leaf keys are no longer stored in memory, reducing churn * key computation is a lot faster thanks to the skipped second disk traversal - a key computation for mainnet can be completed in 11 hours instead of ~2 days (!) thanks to better cache usage and less read amplification - with additional improvements to the on-disk format, we can probably get rid of the initial full traversal method of seeding the key cache on first start after import All in all, this PR reduces the size of a mainnet database from 160gb to 110gb and the peak memory footprint during import by ~1-2gb.
2024-11-20 08:56:27 +00:00
let data = db.mdb.sTab.getOrDefault(rvid, EmptyBlob)
if 0 < data.len:
let key = data.deblobify(HashKey).valueOr:
return err(GetKeyNotFound)
if key.isValid:
return ok(key)
err(GetKeyNotFound)
2024-06-04 15:05:13 +00:00
proc getTuvFn(db: MemBackendRef): GetTuvFn =
result =
2024-06-04 15:05:13 +00:00
proc(): Result[VertexID,AristoError]=
if db.mdb.tUvi.isSome:
return ok db.mdb.tUvi.unsafeGet
err(GetTuvNotFound)
proc getLstFn(db: MemBackendRef): GetLstFn =
result =
proc(): Result[SavedState,AristoError]=
if db.mdb.lSst.isSome:
return ok db.mdb.lSst.unsafeGet
err(GetLstNotFound)
# -------------
proc putBegFn(db: MemBackendRef): PutBegFn =
result =
proc(): Result[PutHdlRef,AristoError] =
ok db.newSession()
proc putVtxFn(db: MemBackendRef): PutVtxFn =
result =
Store keys together with node data (#2849) Currently, computed hash keys are stored in a separate column family with respect to the MPT data they're generated from - this has several disadvantages: * A lot of space is wasted because the lookup key (`RootedVertexID`) is repeated in both tables - this is 30% of the `AriKey` content! * rocksdb must maintain in-memory bloom filters and LRU caches for said keys, doubling its "minimal efficient cache size" * An extra disk traversal must be made to check for existence of cached hash key * Doubles the amount of files on disk due to each column family being its own set of files Here, the two CFs are joined such that both key and data is stored in `AriVtx`. This means: * we save ~30% disk space on repeated lookup keys * we save ~2gb of memory overhead that can be used to cache data instead of indices * we can skip storing hash keys for MPT leaf nodes - these are trivial to compute and waste a lot of space - previously they had to present in the `AriKey` CF to avoid having to look in two tables on the happy path. * There is a small increase in write amplification because when a hash value is updated for a branch node, we must write both key and branch data - previously we would write only the key * There's a small shift in CPU usage - instead of performing lookups in the database, hashes for leaf nodes are (re)-computed on the fly * We can return to slightly smaller on-disk SST files since there's fewer of them, which should reduce disk traffic a bit Internally, there are also other advantages: * when clearing keys, we no longer have to store a zero hash in memory - instead, we deduce staleness of the cached key from the presence of an updated VertexRef - this saves ~1gb of mem overhead during import * hash key cache becomes dedicated to branch keys since leaf keys are no longer stored in memory, reducing churn * key computation is a lot faster thanks to the skipped second disk traversal - a key computation for mainnet can be completed in 11 hours instead of ~2 days (!) thanks to better cache usage and less read amplification - with additional improvements to the on-disk format, we can probably get rid of the initial full traversal method of seeding the key cache on first start after import All in all, this PR reduces the size of a mainnet database from 160gb to 110gb and the peak memory footprint during import by ~1-2gb.
2024-11-20 08:56:27 +00:00
proc(hdl: PutHdlRef; rvid: RootedVertexID; vtx: VertexRef, key: HashKey) =
let hdl = hdl.getSession db
if hdl.error.isNil:
if vtx.isValid:
Store keys together with node data (#2849) Currently, computed hash keys are stored in a separate column family with respect to the MPT data they're generated from - this has several disadvantages: * A lot of space is wasted because the lookup key (`RootedVertexID`) is repeated in both tables - this is 30% of the `AriKey` content! * rocksdb must maintain in-memory bloom filters and LRU caches for said keys, doubling its "minimal efficient cache size" * An extra disk traversal must be made to check for existence of cached hash key * Doubles the amount of files on disk due to each column family being its own set of files Here, the two CFs are joined such that both key and data is stored in `AriVtx`. This means: * we save ~30% disk space on repeated lookup keys * we save ~2gb of memory overhead that can be used to cache data instead of indices * we can skip storing hash keys for MPT leaf nodes - these are trivial to compute and waste a lot of space - previously they had to present in the `AriKey` CF to avoid having to look in two tables on the happy path. * There is a small increase in write amplification because when a hash value is updated for a branch node, we must write both key and branch data - previously we would write only the key * There's a small shift in CPU usage - instead of performing lookups in the database, hashes for leaf nodes are (re)-computed on the fly * We can return to slightly smaller on-disk SST files since there's fewer of them, which should reduce disk traffic a bit Internally, there are also other advantages: * when clearing keys, we no longer have to store a zero hash in memory - instead, we deduce staleness of the cached key from the presence of an updated VertexRef - this saves ~1gb of mem overhead during import * hash key cache becomes dedicated to branch keys since leaf keys are no longer stored in memory, reducing churn * key computation is a lot faster thanks to the skipped second disk traversal - a key computation for mainnet can be completed in 11 hours instead of ~2 days (!) thanks to better cache usage and less read amplification - with additional improvements to the on-disk format, we can probably get rid of the initial full traversal method of seeding the key cache on first start after import All in all, this PR reduces the size of a mainnet database from 160gb to 110gb and the peak memory footprint during import by ~1-2gb.
2024-11-20 08:56:27 +00:00
hdl.sTab[rvid] = vtx.blobify(key)
else:
hdl.sTab[rvid] = EmptyBlob
2024-06-04 15:05:13 +00:00
proc putTuvFn(db: MemBackendRef): PutTuvFn =
result =
2024-06-04 15:05:13 +00:00
proc(hdl: PutHdlRef; vs: VertexID) =
let hdl = hdl.getSession db
if hdl.error.isNil:
2024-06-04 15:05:13 +00:00
hdl.tUvi = some(vs)
proc putLstFn(db: MemBackendRef): PutLstFn =
result =
proc(hdl: PutHdlRef; lst: SavedState) =
let hdl = hdl.getSession db
if hdl.error.isNil:
let rc = lst.blobify # test
if rc.isOk:
hdl.lSst = Opt.some(lst)
else:
hdl.error = TypedPutHdlErrRef(
pfx: AdmPfx,
aid: AdmTabIdLst,
code: rc.error)
proc putEndFn(db: MemBackendRef): PutEndFn =
result =
proc(hdl: PutHdlRef): Result[void,AristoError] =
let hdl = hdl.endSession db
if not hdl.error.isNil:
when extraTraceMessages:
case hdl.error.pfx:
of VtxPfx, KeyPfx: trace logTxt "putEndFn: vtx/key failed",
pfx=hdl.error.pfx, vid=hdl.error.vid, error=hdl.error.code
of AdmPfx: trace logTxt "putEndFn: admin failed",
pfx=AdmPfx, aid=hdl.error.aid.uint64, error=hdl.error.code
of Oops: trace logTxt "putEndFn: failed",
pfx=hdl.error.pfx, error=hdl.error.code
return err(hdl.error.code)
for (vid,data) in hdl.sTab.pairs:
if 0 < data.len:
db.mdb.sTab[vid] = data
else:
db.mdb.sTab.del vid
2024-06-04 15:05:13 +00:00
let tuv = hdl.tUvi.get(otherwise = VertexID(0))
if tuv.isValid:
db.mdb.tUvi = some(tuv)
if hdl.lSst.isSome:
db.mdb.lSst = hdl.lSst
ok()
# -------------
proc closeFn(db: MemBackendRef): CloseFn =
result =
proc(ignore: bool) =
discard
# ------------------------------------------------------------------------------
# Public functions
# ------------------------------------------------------------------------------
proc memoryBackend*(): BackendRef =
let db = MemBackendRef(
beKind: BackendMemory,
mdb: MemDbRef())
db.getVtxFn = getVtxFn db
db.getKeyFn = getKeyFn db
2024-06-04 15:05:13 +00:00
db.getTuvFn = getTuvFn db
db.getLstFn = getLstFn db
db.putBegFn = putBegFn db
db.putVtxFn = putVtxFn db
2024-06-04 15:05:13 +00:00
db.putTuvFn = putTuvFn db
db.putLstFn = putLstFn db
db.putEndFn = putEndFn db
db.closeFn = closeFn db
db
proc dup*(db: MemBackendRef): MemBackendRef =
## Duplicate descriptor shell as needed for API debugging
new result
init_common.init(result[], db[])
result.mdb = db.mdb
# ------------------------------------------------------------------------------
# Public iterators (needs direct backend access)
# ------------------------------------------------------------------------------
iterator walkVtx*(
be: MemBackendRef;
Speed up initial MPT root computation after import (#2788) When `nimbus import` runs, we end up with a database without MPT roots leading to long startup times the first time one is needed. Computing the state root is slow because the on-disk order based on VertexID sorting does not match the trie traversal order and therefore makes lookups inefficent. Here we introduce a helper that speeds up this computation by traversing the trie in on-disk order and computing the trie hashes bottom up instead - even though this leads to some redundant reads of nodes that we cannot yet compute, it's still a net win as leaves and "bottom" branches make up the majority of the database. This PR also addresses a few other sources of inefficiency largely due to the separation of AriKey and AriVtx into their own column families. Each column family is its own LSM tree that produces hundreds of SST filtes - with a limit of 512 open files, rocksdb must keep closing and opening files which leads to expensive metadata reads during random access. When rocksdb makes a lookup, it has to read several layers of files for each lookup. Ribbon filters to skip over files that don't have the requested data but when these filters are not in memory, reading them is slow - this happens in two cases: when opening a file and when the filter has been evicted from the LRU cache. Addressing the open file limit solves one source of inefficiency, but we must also increase the block cache size to deal with this problem. * rocksdb.max_open_files increased to 2048 * per-file size limits increased so that fewer files are created * WAL size increased to avoid partial flushes which lead to small files * rocksdb block cache increased All these increases of course lead to increased memory usage, but at least performance is acceptable - in the future, we'll need to explore options such as joining AriVtx and AriKey and/or reducing the row count (by grouping branch layers under a single vertexid). With this PR, the mainnet state root can be computed in ~8 hours (down from 2-3 days) - not great, but still better. Further, we write all keys to the database, also those that are less than 32 bytes - because the mpt path is part of the input, it is very rare that we actually hit a key like this (about 200k such entries on mainnet), so the code complexity is not worth the benefit really, in the current database layout / design.
2024-10-27 11:08:37 +00:00
kinds = {Branch, Leaf};
): tuple[rvid: RootedVertexID, vtx: VertexRef] =
## Iteration over the vertex sub-table.
for n,rvid in be.mdb.sTab.keys.toSeq.mapIt(it).sorted:
let data = be.mdb.sTab.getOrDefault(rvid, EmptyBlob)
if 0 < data.len:
let rc = data.deblobify VertexRef
if rc.isErr:
when extraTraceMessages:
debug logTxt "walkVtxFn() skip", n, rvid, error=rc.error
else:
Speed up initial MPT root computation after import (#2788) When `nimbus import` runs, we end up with a database without MPT roots leading to long startup times the first time one is needed. Computing the state root is slow because the on-disk order based on VertexID sorting does not match the trie traversal order and therefore makes lookups inefficent. Here we introduce a helper that speeds up this computation by traversing the trie in on-disk order and computing the trie hashes bottom up instead - even though this leads to some redundant reads of nodes that we cannot yet compute, it's still a net win as leaves and "bottom" branches make up the majority of the database. This PR also addresses a few other sources of inefficiency largely due to the separation of AriKey and AriVtx into their own column families. Each column family is its own LSM tree that produces hundreds of SST filtes - with a limit of 512 open files, rocksdb must keep closing and opening files which leads to expensive metadata reads during random access. When rocksdb makes a lookup, it has to read several layers of files for each lookup. Ribbon filters to skip over files that don't have the requested data but when these filters are not in memory, reading them is slow - this happens in two cases: when opening a file and when the filter has been evicted from the LRU cache. Addressing the open file limit solves one source of inefficiency, but we must also increase the block cache size to deal with this problem. * rocksdb.max_open_files increased to 2048 * per-file size limits increased so that fewer files are created * WAL size increased to avoid partial flushes which lead to small files * rocksdb block cache increased All these increases of course lead to increased memory usage, but at least performance is acceptable - in the future, we'll need to explore options such as joining AriVtx and AriKey and/or reducing the row count (by grouping branch layers under a single vertexid). With this PR, the mainnet state root can be computed in ~8 hours (down from 2-3 days) - not great, but still better. Further, we write all keys to the database, also those that are less than 32 bytes - because the mpt path is part of the input, it is very rare that we actually hit a key like this (about 200k such entries on mainnet), so the code complexity is not worth the benefit really, in the current database layout / design.
2024-10-27 11:08:37 +00:00
if rc.value.vType in kinds:
yield (rvid, rc.value)
iterator walkKey*(
be: MemBackendRef;
): tuple[rvid: RootedVertexID, key: HashKey] =
## Iteration over the Markle hash sub-table.
Store keys together with node data (#2849) Currently, computed hash keys are stored in a separate column family with respect to the MPT data they're generated from - this has several disadvantages: * A lot of space is wasted because the lookup key (`RootedVertexID`) is repeated in both tables - this is 30% of the `AriKey` content! * rocksdb must maintain in-memory bloom filters and LRU caches for said keys, doubling its "minimal efficient cache size" * An extra disk traversal must be made to check for existence of cached hash key * Doubles the amount of files on disk due to each column family being its own set of files Here, the two CFs are joined such that both key and data is stored in `AriVtx`. This means: * we save ~30% disk space on repeated lookup keys * we save ~2gb of memory overhead that can be used to cache data instead of indices * we can skip storing hash keys for MPT leaf nodes - these are trivial to compute and waste a lot of space - previously they had to present in the `AriKey` CF to avoid having to look in two tables on the happy path. * There is a small increase in write amplification because when a hash value is updated for a branch node, we must write both key and branch data - previously we would write only the key * There's a small shift in CPU usage - instead of performing lookups in the database, hashes for leaf nodes are (re)-computed on the fly * We can return to slightly smaller on-disk SST files since there's fewer of them, which should reduce disk traffic a bit Internally, there are also other advantages: * when clearing keys, we no longer have to store a zero hash in memory - instead, we deduce staleness of the cached key from the presence of an updated VertexRef - this saves ~1gb of mem overhead during import * hash key cache becomes dedicated to branch keys since leaf keys are no longer stored in memory, reducing churn * key computation is a lot faster thanks to the skipped second disk traversal - a key computation for mainnet can be completed in 11 hours instead of ~2 days (!) thanks to better cache usage and less read amplification - with additional improvements to the on-disk format, we can probably get rid of the initial full traversal method of seeding the key cache on first start after import All in all, this PR reduces the size of a mainnet database from 160gb to 110gb and the peak memory footprint during import by ~1-2gb.
2024-11-20 08:56:27 +00:00
for n,rvid in be.mdb.sTab.keys.toSeq.mapIt(it).sorted:
let data = be.mdb.sTab.getOrDefault(rvid, EmptyBlob)
if 0 < data.len:
let rc = data.deblobify HashKey
if rc.isNone:
when extraTraceMessages:
debug logTxt "walkKeyFn() skip", n, rvid
else:
yield (rvid, rc.value)
# ------------------------------------------------------------------------------
# End
# ------------------------------------------------------------------------------