Skip to content

perf(rpc): hash cached block transactions once - #28552

Merged
mattsse merged 3 commits into
mainfrom
mattsse/rpc-cache-tx-hashes
Nov 14, 2026
Merged

mattsse merged 3 commits into
mainfrom
mattsse/rpc-cache-tx-hashes

Conversation

@mattsse

@mattsse mattsse commented Nov 14, 2026 •

Copy link
Copy Markdown
Collaborator

Transactions read from the database carry no hash. eth_getBlockByNumber/ByHash with full transactions clones the cached block before converting it, so every request hashed every transaction of the block again on the clone, and the cached block never kept the hashes. Keccak was about a third of the node's CPU for these requests.

This bumps reth-core to include paradigmxyz/reth-core#89, where clone_into_rpc_block hashes each transaction through the reference before cloning it. The first full-transaction response fills the hashes on the cached block and later responses reuse them, while callers that never read hashes pay nothing. The bump also brings in the codec changes since the last pin (paradigmxyz/reth-core#72, paradigmxyz/reth-core#81, paradigmxyz/reth-core#83, paradigmxyz/reth-core#84).

Mainnet block 26406938 with full transactions, live node, two interleaved runs per build. Measured at 4950575, which hashed in the state cache instead; not re-measured at the current head, where repeated requests take the same path:

before after
1 connection, requests/s 1060 / 1057 1604 / 1689
1 connection, mean latency 940 µs 605 µs
16 connections, requests/s 9324 / 9251 12472 / 12354

Blocks served from the in-memory chain already carry hashes and are unchanged.

Transactions read from the database carry no hash. A block response with full transactions clones
the cached block before converting it, so every request hashed every transaction of the block
again on the clone, and the cached block never kept the hashes. The state cache now hashes the
transactions of a fetched block once before sharing it. On a mainnet block with full transactions
(eth_getBlockByNumber, block 26406938), one connection went from 1060 to 1650 requests/s
(mean 940 to 605 us) and 16 connections from 9.3k to 12.4k requests/s.
@mattsse
mattsse requested a review from Rjected as a code owner November 14, 2026 06:43
@mattsse mattsse added C-perf A change motivated by improving speed, memory usage or disk footprint A-rpc Related to the RPC implementation labels Nov 14, 2026
@tempoxyz-bot

tempoxyz-bot commented Nov 14, 2026 •

Copy link
Copy Markdown
Member

👁️ Cyclops Security Review

✅ Reviewed 5428905 (latest) · Coverage: partial · no open findings · report
⚠️ Not checked: I could not read the previous pin's derive output and SCALE implementation (reth-core 275ce4cf) as files in this session. The claim that the derive stack path and the inlined SCALE encoder produce the same bytes as the old implementation re
⚠️ Not checked: reth-core bump: RPC full-block conversion via clone_into_rpc_block_full: ruled out without reading the cited code at this commit
⚠️ Not checked: reth-core bump: inlined SCALE Vec decoding of database values: ruled out without reading the cited code at this commit
⚠️ 5 more unchecked questions are listed in the report

Drop the hashing pass from the state cache's block fetch. It ran for
every fetched block, including for callers that never read transaction
hashes (fee history rewards, header and count lookups).

reth-core's `clone_into_rpc_block` now hashes each transaction through
the reference before cloning it, so the first full-transaction response
fills the hashes on the cached block and later responses reuse them.

Bumps reth-core to paradigmxyz/reth-core#89, which also brings in the
codec changes from paradigmxyz/reth-core#72, #81, #83 and #84.
@mattsse
mattsse requested a review from gakonst as a code owner November 14, 2026 08:37
paradigmxyz/reth-core#89 merged as 5dabd396f4, with the same tree as
the PR commit pinned before.
@mattsse
mattsse merged commit 40c24b0 into main Nov 14, 2026
53 checks passed
@mattsse
mattsse deleted the mattsse/rpc-cache-tx-hashes branch November 14, 2026 09:03
opensource-fork-sync Bot pushed a commit to cla7aye15I4nd/opensource-tempoxyz__tempo that referenced this pull request Nov 17, 2026
Automated nightly update of reth dependencies from `paradigmxyz/reth`
main branch.

## Upstream reth changes


[`31d05e1...f551b68`](paradigmxyz/reth@31d05e1...f551b68)

🔗 Amp thread:
https://ampcode.com/threads/T-01a1e993-b347-76cd-8ff2-f35efe2c042f
**RPC**
- Replaced soketto with tokio-tungstenite for WebSockets, then wrote WS
frames directly, without tungstenite. Close replies are now preserved,
and the WS connection closes on oversized fragmented messages
([#28553](paradigmxyz/reth#28553),
[#28557](paradigmxyz/reth#28557),
[#28558](paradigmxyz/reth#28558),
[#28544](paradigmxyz/reth#28544))
- Added timeouts for stalled WebSocket reads and idle HTTP/1 keep-alive
connections ([#28515](paradigmxyz/reth#28515),
[#28545](paradigmxyz/reth#28545))
- HTTP hardening:
- Over-limit streamed request bodies get a 413 response, and refused
upgrades get a 400.
  - Only the request line is logged.
  - The content-type check compares only the JSON media type.
  - Request input is kept out of params and version errors.
- ([#28546](paradigmxyz/reth#28546),
[#28547](paradigmxyz/reth#28547),
[#28493](paradigmxyz/reth#28493),
[#28503](paradigmxyz/reth#28503))
- Server performance work:
  - The RPC service is built once per connection.
  - Sync blocking IO closures run directly.
  - Params are no longer copied for async methods.
  - Response buffers are copied less often.
  - Pubsub notifications are serialized once.
- ([#28522](paradigmxyz/reth#28522),
[#28517](paradigmxyz/reth#28517),
[#28519](paradigmxyz/reth#28519),
[#28500](paradigmxyz/reth#28500),
[#28499](paradigmxyz/reth#28499))
- Faster lookups:
  - Receipt and tx lookup misses take a single provider read.
  - Cached block transactions are hashed once.
  - MEV bundle metadata is no longer cloned per transaction.
- ([#28524](paradigmxyz/reth#28524),
[#28552](paradigmxyz/reth#28552),
[#28512](paradigmxyz/reth#28512))

**Engine / Consensus**
- Split the engine tree handler into modules, and split the chain-state
`in_memory` module so in-memory blocks sit behind a single lock
([#28507](paradigmxyz/reth#28507),
[#28506](paradigmxyz/reth#28506),
[#28543](paradigmxyz/reth#28543))
- BAL hash validation still runs when the requests hash check is
skipped, and BAL retention is capped at the bodies prune distance
([#28535](paradigmxyz/reth#28535),
[#28529](paradigmxyz/reth#28529))
- The engine request span is kept on payload build workers
([#28571](paradigmxyz/reth#28571))

**Trie**
- Queued proof jobs are dropped together with their handle, and V1
proofs receive the sorted overlay
([#28509](paradigmxyz/reth#28509),
[#28570](paradigmxyz/reth#28570))
- Added bucketed fuzz coverage
([#28541](paradigmxyz/reth#28541))

**Storage / DB**
- History reads are floored at the lowest changeset, RPC trie reverts
are no longer cached, and side chain blocks are served by hash
([#28549](paradigmxyz/reth#28549),
[#28511](paradigmxyz/reth#28511),
[#28528](paradigmxyz/reth#28528))
- The CLI archives RocksDB SSTs one file at a time
([#28456](paradigmxyz/reth#28456))

**EVM**
- Absent accounts are added to witness keys
([#28527](paradigmxyz/reth#28527))
- Updated evm2 to e681662
([#28540](paradigmxyz/reth#28540))

**Txpool**
- Added the `--txpool.enforce-tracked-nonce` flag
([#28526](paradigmxyz/reth#28526))

**Build**
- Version metadata is stamped into the binaries, and the build timestamp
is left out of libmdbx
([#28567](paradigmxyz/reth#28567),
[#28573](paradigmxyz/reth#28573))

**Testing**
- Trie workers are drained in the engine hashing task test
([#28554](paradigmxyz/reth#28554))

## Migrations

🔗 Amp thread:
https://ampcode.com/threads/T-01a1e994-010c-74bf-ba4d-8f9ddde55ac8
- Bumped the reth git dependencies (including the `jsonrpsee` entry
pinned to the reth repo) from `31d05e14` to `0e57f259`, `reth-core` from
`275ce4cf` to `5dabd396`, and evm2 and its DaniPopes patch entries from
`66687c18` to `e681662b`.
- Replaced the removed `GasParams::eip7702_auth_state_gas()` helper.
Intrinsic gas now adds `new_account_state_gas()` and
`GasId::TxEip7702PerAuthState` itself.
- Renamed the remaining `eip7702_auth_state_gas()` calls in the
`gas_params.rs` tests to `new_account_state_gas()`. The expected values
(225,000 state, 250,000 total per auth) stay the same.
- Dropped the `state_refund` field from `GasSettlement`, since upstream
removed it. It is gone from `prepare_aa`, `execute_aa`, and every test
that builds a `GasSettlement`.
- `execute_aa` now subtracts the authorization state-gas refund from
`state_gas_spent` itself after `settle_transaction` returns. Before,
`settle_transaction` received it through `state_refund`. The refund goes
back to the reservoir, so it should not count as state gas spent.

[GitHub
Workflow](https://github.com/tempoxyz/tempo/actions/runs/44086786369)

---------

Co-authored-by: Brian Picciano <933154+mediocregopher@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

A-rpc Related to the RPC implementation C-perf A change motivated by improving speed, memory usage or disk footprint

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants