ruby-grape / grape / 35325647120

99%

push

github

Re-bench only master in the version throughput benchmark (#2951)

Every run of run.rb benched every version in DEFAULT_VERSIONS, released
gems included. Their numbers don't move between runs, so a run meant to
measure a change on master spent five sixths of its time -- ~7s per
route shape per JIT mode per version -- measuring rows that were already
known, and brought fresh noise into each of them.

run.rb now keeps every pass that produced a number in
tmp/bench-versions/results.json, under the running Ruby's
RUBY_DESCRIPTION. A plain run benches master plus any default version
with no stored pass for that Ruby -- every version on a first run or a
new Ruby -- and writes RESULTS.md from everything stored. GRAPE_VERSIONS
names exactly the versions to bench again; the report still covers every
default version, sorted by version number rather than by the order they
were listed in. GRAPE_SCENARIOS likewise narrows what is benched without
dropping the other tables. A failed pass is not stored, so its version
is benched again on the next run.

RESULTS.md regenerated that way: master benched on Darwin 27.0.0, the
released rows carried over from the 2026-09-16 report.

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

1424 of 1479 branches covered (96.28%)

Branch coverage included in aggregate %.

4260 of 4260 relevant lines covered (100.0%)

28362.83 hits per line