Exp 162: sustained concurrent reads + pool-cap re-test (paired run)#159
Open
danReynolds wants to merge 1 commit into
Open
Exp 162: sustained concurrent reads + pool-cap re-test (paired run)#159danReynolds wants to merge 1 commit into
danReynolds wants to merge 1 commit into
Conversation
… run) First run under the paired-measurement rule: ships the reads-saturation lane the signal map requested since 2026-04-30 AND executes the implementation it unlocks in the same pass. Measurement: benchmark/profile/sustained_concurrent_reads_audit.dart — 8/16/32 concurrent read clients, no streams, no writes; at the default cap of 4, 12,795/12,800 requests park (max depth = clients - cap) with zero wake retries, satisfying the exp 115 counter gate. Implementation under test: -DRESQLITE_READER_CAP compile-time override (const-folded, tree-shaken at default, not public API) A/Bing caps 4/6/8/12. Reads gain +10-38%, but the exp-147 guardrails reject the raise on the CURRENT dispatch structure: A11c overlap +36%, keyed-PK +29% at cap 8 — exp 105's completion-churn mechanism survives exps 118/120/122, upgrading that rejection from stale to current. Candidate consumed and pruned from signals.json; the lane and counters remain for any future pool-scheduling idea (which must now win on both shapes). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
First experiment executed under the paired-measurement rule (#158): the measurement and the implementation it unlocks, one run, one PR, both results.
Hypothesis
The signal map carried "a long-running concurrent-reads workload that sustains parked dispatchers past pool size" as an open candidate since 2026-04-30. The implementation it unlocks: re-testing the reader-pool cap exp 105 rejected — legitimately re-runnable because exps 118/120/122 restructured dispatch since (the journal's "rejection's reason must have changed" bar).
Approach
benchmark/profile/sustained_concurrent_reads_audit.dart— 8/16/32 concurrent read clients (point + range selects, no streams, no writes — nothing elision or hash suppression can skip), reporting throughput plus the exp 115 dispatcher counters.-DRESQLITE_READER_CAP=ncompile-time override inDatabase.open(const-folded; default builds unchanged; not public API), A/Bing caps 4/6/8/12 on the new lane and 4 vs 8 on the exp-147 guardrails.Results — both halves, per the rule
Measurement (counter gate satisfied): at cap 4, virtually every request parks — 12,795/12,800 at 32 clients, max depth = clients − 4, zero wake retries (exp 118's FIFO holds under sustained overload). The lane the candidate asked for exists and sees what it was meant to see.
Implementation (rejected, with current evidence):
Reads gain +10–38%, but exp 105's completion-churn mechanism survives the 118/120/122 dispatch restructuring — the canonical stream workloads pay far more than the reads lane earns.
Outcome
Measurement accepted; cap raise rejected — rejection upgraded from stale to current. The 2026-04-30 candidate is consumed and pruned. Future pool-scheduling ideas must win on both the reads-saturation lane and the fan-out guardrails; an elastic pool consuming the parked counters is the only shape with a mechanism for that (recorded in signals.json). Single-pass caveat noted in the writeup (one cap-6 outlier omitted); the guardrail deltas are 3–4× today's observed noise bands and confirm a known mechanism.
Test plan
dart analyze lib test benchmarkclean; finalizer greenbenchmark/results/2026-06-11T10-19-38-exp162-…)🤖 Generated with Claude Code