You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Job-minutes cover the CI workflow only, at 100% coverage: 1,515 of 1,519 runs.
The *-compatibility and Benchmarks job data is only partial this run, because the REST quota ran out. From the 1,000 runs sampled (10-01 to 10-02), they add roughly 4–5k job-min/day, about 1.1k of it on macOS. This is mostly from before Skip draft PRs and run macOS legs off the PR path #1093.
In the 24h window there were 287 PR CI runs/day, 105 merge_group runs/day and 35 push runs/day.
History (one row per run, newest first, up to 18 rows)
Related open ci-janitor work, not counted above: #1143 (decouple e2e-build per OS, about 5 min of merge-queue critical path), #1146 (fail ci-ok early on doomed queue entries, about 35 min per failed entry). #1149 recommended a 120-min merge-queue check timeout; ruleset 24668460 already shows check_response_timeout_minutes: 120.
Merged ci-perf / ci-janitor PRs, last 7 days: measured effect
PR
Merged
Effect, before → after
#1093 Skip draft PRs; run macOS legs off the PR path
10-08 02:16
✅ macOS per PR CI run went from 76.2 to 4.1 job-min. macOS queue p50 went from 45–88 min (10-07 18Z–22Z) to ≤0.8 min from 04Z on; p90 since 04Z is 2.0 min.
#1133 Shard Gradle e2e and test legs; skip test-release in the queue
10-08 18:55
✅ merge_group run → ci-ok p50 went from 40.8 min (10-08 02:16–18:55) to 22–25 min (n=5 complete runs so far).
Closing as a duplicate of #1182. Two profiler runs each created a dashboard. The history rows from this issue were merged into #1182, which is now the single CI performance dashboard.
Maintained by the CI profiler routine. It runs every 4 hours and replaces this body each time. Last run: 2026-10-08 19:36Z, the first run.
Headline (windows: 7d = 10-01 19:36Z → 10-08 19:36Z; 24h = the last 24h)
ci-ok, p50 / p90ci-ok, p50 / p90Notes:
CIworkflow only, at 100% coverage: 1,515 of 1,519 runs.*-compatibilityandBenchmarksjob data is only partial this run, because the REST quota ran out. From the 1,000 runs sampled (10-01 to 10-02), they add roughly 4–5k job-min/day, about 1.1k of it on macOS. This is mostly from before Skip draft PRs and run macOS legs off the PR path #1093.History (one row per run, newest first, up to 18 rows)
Open
ci-perfissues, ranked by ROIROI = weighted saving × confidence ÷ effort.
llvm-cov reporttwiceRelated open
ci-janitorwork, not counted above: #1143 (decouple e2e-build per OS, about 5 min of merge-queue critical path), #1146 (failci-okearly on doomed queue entries, about 35 min per failed entry). #1149 recommended a 120-min merge-queue check timeout; ruleset 24668460 already showscheck_response_timeout_minutes: 120.Merged
ci-perf/ci-janitorPRs, last 7 days: measured effectci-okp50 went from 40.8 min (10-08 02:16–18:55) to 22–25 min (n=5 complete runs so far).test (ubuntu-latest)leg is gone: 433 legs before, 0 after, at about 11 min each.No merged fix has been found to under-deliver so far.