Experiment 097: One-pass initial stream decode and hash

Date: 2026-04-23

Status: In Review

Problem

Initial stream registration decodes the query result for subscribers, then

replays the same SQLite statement through resqlite_query_hash to establish

the baseline result hash used by later selectIfChanged calls.

That two-pass setup is correct, but it is wasteful for workloads that create

many streams or frequently subscribe/cancel.

If the initial stream path fills the existing cell buffer and folds the native

FNV result hash during the same SQLite step pass, stream setup should improve

without changing public API behavior or the unchanged-result fast path.

Added a native resqlite_step_row_hash helper used only by initial stream

registration. It mirrors resqlite_step_row, but also updates the same masked

FNV accumulator used by resqlite_query_hash.

Dart now uses a dedicated decodeQueryWithInitialHash path for

selectWithDeps. Normal select() still uses the existing decoder, and later

stream re-queries still use the hash-only resqlite_query_hash path so

unchanged results can skip Dart decoding entirely.

Artifacts:

benchmark/results/2026-04-23T19-38-11-exp097-one-pass-initial-stream-hash.md
benchmark/results/2026-04-23T19-38-11-exp097-one-pass-initial-stream-hash.json

Release benchmark summary: 6 wins, 0 regressions, 147 neutral.

Key stream metrics:

Benchmark	Previous	Current	Result
Fan-out (10 streams)	0.24 ms	0.21 ms	14% faster
Stream subscription rate (500 subscribe+cancel)	7.22 ms	6.06 ms	16% faster
Stream churn (100 cycles)	2.06 ms	1.53 ms	faster, but below threshold
Initial emission	0.03 ms	0.03 ms	tiny absolute delta, within noise

The structured artifact also showed the expected direction on the targeted

setup-heavy paths:

Benchmark	Baseline artifact	Experiment artifact
Initial emission	0.036 ms	0.025 ms
Fan-out (10 streams)	0.267 ms	0.211 ms
Stream churn (100 cycles)	1.898 ms	1.535 ms
Stream subscription rate (500 cycles)	7.646 ms	6.063 ms

Correctness checks passed:

dart test test/stream_test.dart test/stream_invalidation_coalescing_test.dart test/stream_dependency_shapes_test.dart test/reader_pool_test.dart

Keep in review.

This is the only candidate from the sweep with a clean, targeted signal and no

reported release regressions. The main trade-off is code complexity: the

initial stream path now has a second decode loop that must stay hash-compatible

with resqlite_query_hash. The win is concentrated exactly where expected,

which makes it worth carrying to PR review.