Skip to content

exp 184: writer residual re-split (post-159/180) — measurement - #186

Merged
danReynolds merged 1 commit into
mainfrom
exp-184-writer-residual-resplit
Jun 17, 2026
Merged

exp 184: writer residual re-split (post-159/180) — measurement#186
danReynolds merged 1 commit into
mainfrom
exp-184-writer-residual-resplit

Conversation

@danReynolds

Copy link
Copy Markdown
Owner

Measurement experiment (no runtime code → auto-merge eligible).

Re-ran exp 147's writer_sqlite_wall_audit on main after pipelining (exp 159) + group commit (exp 180) to re-rank the next writer target.

workload (sequential) SQLite invalidation residual exp 147
A11c baseline 38% 0% 62%
A11c disjoint 24% 22% 55% ~55%
A11c overlap 14% 16% 70% 72%
keyed PK 24% 14% 62% 63%

Essentially unchanged — expected: the audit issues writes sequentially (audit_workloads.dart:167) and exp 180 coalesces only concurrent bursts, so the sequential path it measures is unmoved (exp 180's win is on the concurrent release lane, validated separately).

Conclusion — retire writer-residual micro-opts:

  • Residual (55–70%) = the per-write isolate round-trip = SDK-gated floor (lever: shared-memory transport).
  • Invalidation/re-query precision (14–22%) = owned by in-flight exp 160 (Exp 160: incremental view maintenance — the v2 stream engine #155); traversal already at floor (exp 121).
  • Dirty harvest (~4–6%, exp 182) = small; gating it regresses with-streams.
  • SQLite (14–38%) = sqlite3mc config, mined.

Next reactive-engine win = exp 160; next sequential-write reduction = shared-memory transport.

🤖 Generated with Claude Code

Re-ran exp 147's writer_sqlite_wall_audit on main after pipelining (159) +
group commit (180). Breakdown essentially unchanged (A11c overlap SQLite 13.6%
/ invalidation 16.0% / residual 70.3% vs exp 147's 9.4/18.8/71.8): the audit
issues writes sequentially and exp 180 coalesces only concurrent bursts, so the
sequential writer path it measures is unmoved.

Retires writer-residual micro-opts from the candidate list: the residual is the
SDK-gated per-write round-trip; invalidation precision (14-22%) is owned by
in-flight exp 160 (#155) with its traversal at floor (exp 121); dirty harvest
(~4-6%, exp 182) is small; SQLite (14-38%) is mined. Next reactive-engine win =
exp 160; next sequential-write reduction = shared-memory transport.

No runtime code; profile-mode measurement only (Benchmark Run: none).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@danReynolds
danReynolds merged commit 5fbce43 into main Jun 17, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant