---

pgmcp_experiment: parallel-batch-fuzzy-query-evaluation-over-a-shared-transducer title: Parallel batch fuzzy-query evaluation over a shared transducer date: 2026-07-05 project: liblevenshtein-rust kind: optimization status: decided verdict: accepted p_value: 0.000000 git_ref: cd385e3

Parallel batch fuzzy-query evaluation over a shared transducer

Kind: optimization | Status: decided | Correction: benjamini_hochberg

Method

Question: Phase 6 novel option #3 (specialize for observed query distributions) / SOTA parallel batch evaluation: the Transducer is Send+Sync (PathMapDictionary = Arc), so a batch (distribution) of independent fuzzy queries can be evaluated with rayon over a shared read-only transducer. Does parallel batch evaluation reduce whole-batch wall time versus sequential on a 512-query batch?

Transducer holds dictionary: D; PathMapDictionary is Arc-backed (Sync). query() takes &self, so a shared transducer can be queried across threads.

Hypotheses

H1. Evaluating a 512-query batch with rayon par_iter over a shared &Transducer reduces whole-batch wall time versus sequential iteration by a large margin (multi-core, embarrassingly parallel), returning identical per-query result counts. — ✅ accepted

• metric: batch_wall_time_ms (ms) · predicted: decrease · planned n/arm: 30
• pre-registered criterion (locked 2026-07-05 00:42:32Z): {"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
batch_wall_time_ms	welch_t	-3663.538190	0.000000	-529.084858	[-161.9787, -161.8040]	accepted

Decision on batch_wall_time_ms:

ACCEPTED (criterion: welch_t, correction: BenjaminiHochberg) [0] WelchT: statistic=-3663.5382, p=0.000000, effect=-529.0849, 95% CI=[-161.9787, -161.8040]

Operator note: Parallel batch fuzzy-query evaluation (512-query batch, dict 2000, d=2) via rayon par_iter over a shared read-only &Transducer, pinned to one 8-core CCD. Control (sequential) 185.98 ms vs treatment (parallel) 23.81 ms — ~7.8x on 8 cores (near-linear). Correctness gate passed: parallel per-query result counts identical to sequential. Transducer is Send+Sync (PathMapDictionary = Arc), so no code change needed — the win is an opt-in usage pattern (par_iter over queries).

What did NOT work

Nothing rejected (or no decisions yet).

Reproducibility

• git ref: cd385e3
• See each hypothesis's pre-registered criterion above; raw samples are retained in experiment_samples.

Timeline

• 2026-07-05 00:42:32Z — opened: Parallel batch fuzzy-query evaluation over a shared transducer
• 2026-07-05 00:42:32Z — criterion_locked: Evaluating a 512-query batch with rayon par_iter over a shared &Transducer reduces whole-batch wall time versus sequential iteration by a large margin (multi-core, embarrassingly parallel), returning identical per-query result counts.
• 2026-07-05 00:44:54Z — run: treatment (treatment)
• 2026-07-05 00:46:00Z — run: control (control)
• 2026-07-05 00:46:13Z — decided: accepted on batch_wall_time_ms (welch_t)

---

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