Analysis methodology
The analysis package (analyze-benchmarks) never rewrites raw logs.
- Load the latest
logs/<lang>/*.csv(or an explicit stem). - Drop warmup when
statistics.exclude_warmupis true (index< warmup_repetitions). - Outliers: IQR with
k=1.5(configurable). Raw rows stay in the CSV. - Summaries: mean, median, std, MAD, CV, percentiles, min/max.
- Bootstrap 95% CI on the mean (percentile, 2000 iterations, seed 42).
- Rank by
total_median_ns(handoff). Lower is better. - Effect sizes vs the fastest in the group (Cliff’s δ, Hedges’ g).
- Regression gate: fail only when the point estimate and the optimistic
CI end are both more than
threshold_percentslower than baseline.
Compare queues inside one language and one communication category (T or A). See Benchmark design.