fix(trie): start partial proofs below known parent by mediocregopher · Pull Request #26471 · paradigmxyz/reth

mediocregopher · 2026-07-22T08:33:41Z

Start known-parent proofs at direct children and batch targets sharing a parent into one bounded traversal. This prevents a stale masked parent node (due to partial persistence) from suppressing live child ranges without splitting sibling targets into independent cursor traversals (which is bad for perf).

For targets aed2 and aed4 under known parent aed:

Before (origin/main)

Seek aed and traverse [aed, aee), putting the parent and all 16 child ranges in scope.
The database parent can be consulted even though the sparse trie already supplied the current parent.

After

Seek the first targeted child, aed2, and traverse the bounded sibling span [aed2, aed5) once. The parent is never read by the trie traversal, while outer siblings aed0–aed1 and aed5–aedf are excluded from calculation.
Targeted children aed2 and aed4 are calculated and retained independently.
Untargeted aed3 stays inside the traversal span but is not retained. Usable cached trie data skips its leaves; if it is uncached, the hashed cursor advances through it sequentially instead of performing another seek at aed4.

This narrows the work from the whole parent range while preserving efficient sequential traversal between targeted siblings.

Testing

Ran 1k big blocks, 20k mainnet blocks, and 10k mainnet blocks with reorgs with no issues.

# Conflicts: # crates/e2e-test-utils/tests/rocksdb/main.rs # crates/engine/primitives/src/config.rs # crates/engine/tree/src/tree/state_root_strategy/mod.rs # crates/node/core/src/args/engine.rs

mediocregopher · 2026-07-22T09:10:16Z

derek bench

mediocregopher · 2026-07-22T09:10:26Z

derek bench big-blocks bal=true

decofe · 2026-07-22T09:10:32Z

cc @mediocregopher

✅ Benchmark complete! View run

Benchmark Results

Configuration

Derek command: derek bench blocks=500 warmup=125 baseline=main feature=mediocregopher/proof-v2-child-subtrie bal=false cores=0 run-pairs=6 otlp=true metrics=false slack=always

⚠️ Feature is 1 commit behind main. Consider rebasing for accurate results.

Metric	`main`	`mediocregopher/proof-v2-child-subtrie`	Change
Mean	34.68ms	34.37ms	-0.89% ⚪ (±0.52%, floor 1.20%)
P50	24.09ms	23.74ms	-1.46% ⚪ (±0.51%, floor 1.20%)
P90	73.62ms	72.97ms	-0.88% ⚪ (±2.32%, floor 1.35%)
P99	188.49ms	193.68ms	+2.75% ⚪ (±5.02%, floor 5.00%)
Mgas/s	1146.22	1160.13	+1.21% ⚪ (±0.52%, floor 1.20%)
Wall Clock	105.00s	104.01s	-0.94% ⚪ (±0.57%, floor 0.70%)
Persist Wait	22.26ms	22.51ms	+1.12% ⚪ (±1.16%, floor 5.00%)

500 blocks, mode: engine, 125 warmup, 6 run pairs

Wait Time Breakdown

Persistence Wait

Metric	`main`	`mediocregopher/proof-v2-child-subtrie`
Mean	22.26ms	22.51ms
P50	0.02ms	0.02ms
P95	227.00ms	224.74ms

Trie Cache Update Wait

Metric	`main`	`mediocregopher/proof-v2-child-subtrie`
Mean	0.00ms	0.00ms
P50	0.00ms	0.00ms
P95	0.00ms	0.00ms

Execution Cache Update Wait

Metric	`main`	`mediocregopher/proof-v2-child-subtrie`
Mean	0.00ms	0.00ms
P50	0.00ms	0.00ms
P95	0.00ms	0.00ms

Observability

Logs
Traces

Charts

Latency, Throughput & Diff

Wait Time Breakdown

Gas vs Latency

decofe · 2026-07-22T09:10:40Z

cc @mediocregopher

✅ Benchmark complete! View run

Benchmark Results