Repository navigation
CI performance dashboard #1175
Description
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)
| Metric | 24h | 7d |
|---|---|---|
| PRs merged | 43 | 204 (29/day) |
| Merge queue, last enqueue → merge, p50 / p90 | 83 / 177 min | queue live since about 10-07 20:35Z (n=43) |
| Merge queue, first enqueue → merge, p50 / p90 | 128 / 673 min | 26 re-enqueues (12 PRs evicted at least once) |
merge_group run → ci-ok, p50 / p90 |
42.6 / 70.2 min | after #1133 (18:55Z): 22–25 min |
PR push → ci-ok, p50 / p90 |
41.6 / 78.5 min | 42.7 / 147.9 min |
| macOS queue wait, p50 / p90 | 1.1 / 69.8 min. Since 10-08 04Z: 0.2 / 2.0 | 0.1 / 36.9 min |
| Linux queue wait, p50 / p90 | 0.1 / 7.6 min. Busy hours: p50 5–10 min | 0.1 / 6.0 min |
| Windows queue wait, p50 / p90 | 0.1 / 0.4 min | 0.1 / 0.3 min |
| CI job-min/day: Linux / macOS / Windows | 139k / 21k / 21k | 50k / 12.6k / 8.8k |
Notes:
- Job-minutes cover the
CIworkflow only, at 100% coverage: 1,515 of 1,519 runs. - The
*-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. - 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)
| Run (UTC) | Merges/24h | MQ enqueue→merge p50/p90 (min) | merge_group→ci-ok p50 (min) | PR push→ci-ok p50/p90 (min) | macOS queue p50/p90 (min) | Linux queue p50/p90 (min) | Job-min/day L/M/W (24h, CI) |
|---|---|---|---|---|---|---|---|
| 2026-10-08 19:36 | 43 | 83 / 177 | 42.6 (22–25 after #1133) | 41.6 / 78.5 | 1.1 / 69.8 | 0.1 / 7.6 | 139k / 21k / 21k |
Open ci-perf issues, ranked by ROI
ROI = weighted saving × confidence ÷ effort.
- Weighted saving counts thousands of job-min/day as Linux ×1, Windows ×2 and macOS ×3, plus 1.0 per critical-path minute.
- Effort is S=1, M=2, L=3.
| # | Issue | ROI | Saving |
|---|---|---|---|
| 1 | #1170 CI on push to main re-tests the SHA the merge queue already passed | 8.4 | about 12.2k Linux, 1.8k macOS and 1.7k Windows job-min/day |
| 2 | #1171 Gradle e2e shards unbalanced, which puts them on the merge-queue critical path | 2.1 | about 3 min per merge_group run and on PR latency |
| 3 | #1172 e2e fan-out: 82–87% of legs run under 2 min | 1.9 | about 6k Linux and 1.2k macOS job-min/day |
| 4 | #1173 test (windows/macos) has a redundant Build step and full debuginfo | 1.3 | about 630 Windows and 300 macOS job-min/day |
| 5 | #1174 coverage runs llvm-cov report twice |
0.42 | about 530 Linux job-min/day |
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). |
| #1139 Fetch Maven from Central | 10-08 18:55 | Too early to measure. Will re-check next run. |
| #1080 Stop cancelling in-progress Composer runs on main | 10-08 00:33 | Not a perf change. |
| #619 Stop running the Linux test suite twice | 10-05 11:18 | ✅ the test (ubuntu-latest) leg is gone: 433 legs before, 0 after, at about 11 min each. |
| #897 / #892 / #874 / #500 (pnpm per-Node jobs, cancel superseded compat runs, macOS vexctl cache, vexctl download) | 10-02 to 10-07 | Not measured this run, because compatibility-workflow job data is partial. |
No merged fix has been found to under-deliver so far.
Activity
mikolalysenko commentedon Oct 8, 2026
[agent] Triaged as
priority:p3. This is a routine-maintained tracking dashboard, not an actionable defect, so agents won't claim it.Generated by Claude Code
mikolalysenko commentedon Oct 9, 2026
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.Generated by Claude Code