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
{{ message }}
Repository navigation
Lower merge-queue check timeout from 360 to 120 min #1149
The merge-queue check timeout in ruleset 24668460 was raised from 90 to 360 minutes as a stopgap on 2026-10-08. The data supports lowering it to 120 minutes.
Numbers (CI workflow, merge_group event, 2026-10-01 to 2026-10-08, run start to last update):
45 successful runs: p50 47 min, p90 81 min, max 111 min. A 120-minute timeout would have evicted none of them.
Why it matters: the check timeout only applies when a run hangs or never reports. At 360 min, a stuck ci-ok holds the queue head, and the up-to-5 entries built on it, for up to 6 hours before eviction. At 120 min it's 2 hours, and no healthy run in the last week came near that.
Related:#1146 makes a doomed merge-group run fail ci-ok early instead of waiting for the slowest e2e leg. #1133 shards the merge-group critical path (~46 min to ~20 min). Once both land, 90 min would again leave about 4x headroom over p50.
Recommended: check timeout 120 min now; 90 min after #1133 lands. Keep the build concurrency at 5.
[agent] Triage: priority:p3 (CI only). The fix is a change to repository ruleset 24668460 (the merge-queue check timeout). That is an admin setting, not something in the repo, so no code PR can make it. Labeled agent:needs-human for a maintainer with admin access to apply.
Done. Ruleset 24668460's merge-queue check_response_timeout_minutes is now 90. It was already lowered from 360 to 120 on 2026-10-08.
The 90-minute step applies now because #1133 (merged 2026-10-08 18:55Z) and #1146 (21:05Z) have both landed. Since #1133, the last 40 successful merge_group CI runs had p50 22 min, p90 24 min and max 28 min, so 90 min is about 4x p50 and about 3x the slowest run.
Note: max_entries_to_build is currently 8, not the 5 this issue recommended. That was changed separately, and I left it as is.
The merge-queue check timeout in ruleset 24668460 was raised from 90 to 360 minutes as a stopgap on 2026-10-08. The data supports lowering it to 120 minutes.
Numbers (
CIworkflow,merge_groupevent, 2026-10-01 to 2026-10-08, run start to last update):rewrite()lost an argument) failed every entry built on Fix open bun issues (#992, #861, #784, #764, #735, #635, #599, #578, #497, #443, #371) #1009. That was 6 runs at 15:14 UTC (all four test/coverage jobs failed at about 2-4 min). The queue was then rebuilt twice: 1117 at 15:33, and 1110/1106/1103/1091 at about 15:55, cancelled at 16:29. Entries enqueued before 15:14 were still in a fresh batch at 16:37.Why it matters: the check timeout only applies when a run hangs or never reports. At 360 min, a stuck
ci-okholds the queue head, and the up-to-5 entries built on it, for up to 6 hours before eviction. At 120 min it's 2 hours, and no healthy run in the last week came near that.Related: #1146 makes a doomed merge-group run fail
ci-okearly instead of waiting for the slowest e2e leg. #1133 shards the merge-group critical path (~46 min to ~20 min). Once both land, 90 min would again leave about 4x headroom over p50.Recommended: check timeout 120 min now; 90 min after #1133 lands. Keep the build concurrency at 5.
Generated by Claude Code