| name | rewrite-hashes |
| description | Use when a dashboard benchmark series ends or splits while the benchmark still runs — after newton renames a benchmark or changes its source (asv version hash), or when check-regressions flags hash mismatches or disappeared benchmarks. |
Rewrite Benchmark Hashes
Overview
asv matches results to a dashboard series by benchmark key + version hash (hash of the benchmark source). When newton changes a benchmark, old result files keep the old hash and the dashboard starts a new series. results/adenzler-asv-runner/replace_hash.sh holds a name→current-hash map (must match results/benchmarks.json) and rewrites the hashes inside all result JSONs to reconnect history.
Prerequisite: at least one post-change automation run has landed, so results/benchmarks.json already contains the new hashes.
Decision gate — before touching anything
Establish the facts per benchmark:
- What changed?
git -C <newton> log -p <range> -- asv/ for the benchmark file AND the example it wraps (check imports in newton's asv/benchmarks/), where <newton> is a local checkout of github.com/newton-physics/newton.
- Is there a step at the series boundary? (
series.py from the check-regressions skill — note the magnitude per machine.)
Timing-irrelevant change (pure rename, comments, untimed setup) with no boundary step → merge (procedures below).
Anything else — workload, timed region, or scene changed, or a boundary step exists — is a product decision with no clean rule: continuous history is often worth more than strict comparability, so "the setup changed" does not imply "don't merge". Never decide this yourself. Present the evidence (what changed, step magnitude per machine) and ask the user, offering:
- Merge: history stays continuous; the change appears as a known, documented step in the series. Often right when it's still "the same benchmark", measured somewhat differently.
- Don't merge: old series ends, new baseline starts; drop/skip the replace_hash.sh entry with a comment documenting the exclusion (precedent: tiled-camera #3480, new scene).
- Defer: leave as-is until more post-change data exists to judge the step.
Raw old data stays in the result files either way; whichever way it goes, record the decision in the commit message and a script comment so it isn't re-litigated.
Never add a benchmark to replace_hash.sh just because an audit finds it "missing" — absence may be a deliberate exclusion, and the reflex once silently rewrote 109 result files across a timed-region change.
Procedures
| Case | What to do | Precedent commit |
|---|
| Same name, new hash | Update the entry in replace_hash.sh from benchmarks.json; run ./replace_hash.sh <path-to-results> | most history of the script |
| Renamed, hash unchanged | Rename the key in all historical result JSONs; update script entries | 1cd00414 |
| Renamed, old+new keys coexist in files | Merge rows: union the param combos so retired values stay under their old params | 292e16b9 |
git show <precedent> shows the exact mechanics for the rename/merge cases.
Verify and land
git diff --stat: only intended files. Spot-check a hunk: only 64-hex version strings (or renamed keys) changed.- Re-dump one affected series: continuous, no artificial boundary step.
- Convention: these maintenance commits land directly on main (data repo). Remote pushes still need explicit user approval.
Common mistakes
- Running before benchmarks.json has the new hashes — rewrites history to stale values.
- Forgetting the script defaults to its own directory — pass the results root explicitly.
- Treating replace_hash.sh as an exhaustive registry. It is a merge tool; a benchmark not listed is a decision, not an oversight.
