---

pgmcp_experiment: version-tied-cross-query-result-cache-for-the-levenshtein-transducer title: Version-tied cross-query result cache for the Levenshtein transducer date: 2026-07-05 project: liblevenshtein-rust kind: optimization status: decided verdict: accepted p_value: 0.000000 git_ref: 82cce83

Version-tied cross-query result cache for the Levenshtein transducer

Kind: optimization | Status: decided | Correction: benjamini_hochberg

Method

Question: AMD uProf (tbp) shows transducer::transition::transition_state_pooled_ref dominates query CPU at ~80% (100.9s). Does a dictionary-version-tied query-result cache that skips the entire transition loop on repeated queries reduce mean per-query latency on a repeated-query workload, without regressing unique-query workloads unacceptably?

Phase 6 novel option #4 (cross-query cache with eviction tied to dictionary versioning). Profiled with AMD uProf tbp; transition_state_pooled_ref is the sole hot function. A cache hit skips it entirely.

Hypotheses

H1. A VersionedQueryCache memoizing (query, max_distance) -> results and invalidating on a dictionary version bump reduces mean per-query latency on a repeated-query workload by a large margin versus the uncached transducer, while returning identical results. — ✅ accepted

• metric: mean_query_latency_ns (ns) · predicted: decrease · planned n/arm: 30
• pre-registered criterion (locked 2026-07-05 00:13:14Z): {"type": "welch_t", "params": {"tail": "less", "alpha": 0.05, "min_effect": {"kind": "cohens_d", "threshold": 0.5}}}

Measurements & Decisions

Metric	Test	Statistic	p	Effect	95% CI	Verdict
mean_query_latency_ns	welch_t	-619.411566	0.000000	-80.866202	[-285238.0267, -283414.9544]	accepted

Decision on mean_query_latency_ns:

ACCEPTED (criterion: welch_t, correction: BenjaminiHochberg) [0] WelchT: statistic=-619.4116, p=0.000000, effect=-80.8662, 95% CI=[-285238.0267, -283414.9544]

Operator note: Version-tied cross-query cache vs uncached transducer on a repeated-query workload (50 queries × replicates, dict 1000 terms, max_distance 2), CPU-pinned core 2. Cache hits skip the AMD-uProf-dominant transition_state_pooled_ref (~80% CPU). Correctness gate passed (cached results identical to uncached for all 50 queries).

What did NOT work

Nothing rejected (or no decisions yet).

Reproducibility

• git ref: 82cce83
• See each hypothesis's pre-registered criterion above; raw samples are retained in experiment_samples.

Timeline

• 2026-07-05 00:13:14Z — opened: Version-tied cross-query result cache for the Levenshtein transducer
• 2026-07-05 00:13:14Z — criterion_locked: A VersionedQueryCache memoizing (query, max_distance) -> results and invalidating on a dictionary version bump reduces mean per-query latency on a repeated-query workload by a large margin versus the uncached transducer, while returning identical results.
• 2026-07-05 00:19:55Z — run: control (control)
• 2026-07-05 00:21:21Z — run: treatment (treatment)
• 2026-07-05 00:21:48Z — decided: accepted on mean_query_latency_ns (welch_t)

---

Rendered from the pgmcp experiment record (the structured source of truth). Edit the experiment, not this file.