PR benchmark regression check

Compares a PR's benchmark run against its baseline (merge-base on next, optionally a window of baseline runs) and reports only the regressions that (a) are statistically real and (b) plausibly relate to what the PR changed. The data-pulling and diff math are done by ci3/bench_compare; your job is to get a run, invoke the script, then filter the output with judgement.

How benchmark data is stored (context)

Each CI benchmark run uploads a single bench-out/bench.json (a flat array of {name, unit, value}, customSmallerIsBetter — higher = worse) to the build-cache, keyed by the commit's git tree hash (bench-<treehash>.tar.gz, ~40 KB). A PR only produces one on an uploadable run: labels ci-full-no-test-cache, ci-full, or a merge-queue run. ci3/bench_compare pulls these per-run blobs — never the multi-MB per-branch data.js graph history.

Steps

1. Get a benchmark run for the PR

• Default to HEAD (or the ref the user gives). Try step 2 directly — ci3/bench_compare fails clearly if the commit has no bench data.
• If there's no run and the user is fine using the last one: walk back the branch for the most recent benched commit — for c in $(git rev-list -n 30 HEAD); do ci3/bench_compare "$c" >/dev/null 2>&1 && echo "$c" && break; done — and use that commit.
• If a fresh run is needed: tell the user to add the ci-full-no-test-cache label to the PR (that triggers x-bench on a dedicated metal box and uploads) and to re-run this once CI finishes. Do not block waiting unless asked.

2. Determine the baseline branch — do NOT assume next

The comparison is merge-base(PR, baseline), so the baseline MUST be the PR's actual target branch. A PR onto v5-next or a merge-train/* branch compared against next would report the entire branch-vs-branch perf delta as bogus regressions. Get it from the PR itself:

gh pr view <pr-or-branch> --json baseRefName -q .baseRefName   # e.g. next, v5-next, merge-train/spartan
git fetch origin <baseRefName>                                  # ensure origin/<baseRefName> is current

(You do NOT need to worry about cross-branch data mixups: the bench cache is content-addressed by git tree hash, so next, v5-next, etc. already occupy different keys — a next commit's tree is only ever benched by next-content. The only thing you control, and must get right, is which branch's lineage you merge-base against.)

3. Run the comparison

ci3/bench_compare <pr-commit> --baseline origin/<baseRefName> --window 5 --out /tmp/bench-report.json

• Always pass --baseline explicitly (the script auto-detects via gh if omitted, but be explicit). It prints baseline = … to stderr — confirm it matches the PR's target before trusting the numbers.
• --window 5 averages the merge-base + 4 preceding baseline runs, yielding stddev and z per bench — essential for separating real regressions from flaky benches. Use --window 1 if history is sparse.
• JSON on --out is { meta, benches:[{name,unit,pr,baseline,n,pct,stddev,z,status}] }, sorted by pct desc. Human summary on stderr.

4. Filter to what matters (your judgement)

Read the JSON and keep a regression only if all hold:

• Statistically real: z >= ~4 (the PR value is ≥4 baseline-stddevs above the mean). A large pct with small/absent z is baseline noise — a flaky bench (e.g. p2p/BatchTxRequester/.../duration hitting a 30 s timeout, pct in the thousands) — ignore it. When window is 1 (no z), fall back to requiring a large pct AND a meaningful absolute magnitude.
• Material magnitude: skip tiny absolutes (e.g. 0.01 ms -> 0.02 ms = +100% but noise).
• Plausibly caused by the PR: map the bench name prefix to the PR's changed areas. Get the diff with git diff --name-only $(git merge-base HEAD origin/next)..HEAD | cut -d/ -f1-2 | sort -u, then keep benches whose names start with a touched area (yarn-project/simulator, barretenberg/cpp, avm-transpiler, …) and treat far-away regressions (a TS-only PR moving a C++ proving bench) as drift/noise, not this PR.

5. Report

List the surviving regressions with baseline -> pr unit, pct, and z, grouped by area, and give a one-line verdict (clean / N real regressions in <area>). Note anything you filtered out and why, so the dev can override your judgement.

Notes

• Baseline is origin/next by default; git fetch it first if stale (ci3/bench_compare errors if the ref is missing).
• Requires jq, python3, and read access to the build-cache (the public build-cache.aztec-labs.com endpoint needs no creds; the in-VPC S3 path is a fallback).
• To author or wire up new benchmarks, see the adding-benchmarks skill; this skill only checks them.