CI perf (settings): merge queue builds 5 entries at a time — a 13-PR burst took 100 min to drain on a 21-min CI run (~21 min per PR past the 5th) #1248

Filed by the CI profiler routine (observer only). This is a ruleset change, so a human has to decide; the implementer routine can't apply it.

Measurement

Window: 2026-10-08 08:17 → 2026-10-09 08:17 UTC. Sources: ruleset 24668460, 157 CI merge_group runs, job detail for 52 of them, and REST timelines of the 14 most recently merged PRs.

Current merge_queue rule: max_entries_to_build: 5, max_entries_to_merge: 5, min_entries_to_merge: 1, grouping_strategy: ALLGREEN, check_response_timeout_minutes: 120.

What happened in the 06:19–08:01 burst. 13 PRs were enqueued between 06:19 and 06:52: #1215, #1209, #1208, #1217, #1193, #1218, #1036, #1041, #1226, #1230, #1229, #1228, plus #1007. They were tested in waves of at most 5 CI runs, and each wave started only after the previous one finished:

So the queue's throughput is 5 PRs per ~21 min (≈14/h). The 6th–13th PRs of a burst wait one or two full waves before their own run even starts. Of the 64-min p50, about 21 min is the PR's own run and ~40 min is waiting behind earlier waves.

Where the time goes

• Per wave: test (windows-latest, 2) (Run tests 15.0 min p50 + Build 2.2), coverage-docker (sbt) → coverage-merge (CI perf: coverage-docker (sbt) — 5.6-min image rebuild warms 3 unused toolchains; merge-queue critical path (~2.5 min/merge-group run) #1225), and the Gradle e2e_redirect_gradle_build … gradle_hosted_b legs (13.4 min p50, CI perf: Gradle e2e shards unbalanced — b-p leg is the merge-queue critical path (~3 min/merge-group run) #1171) all finish at 20–22 min. Queue wait for those jobs is ≤0.1 min (Linux/Windows) and 0.2 min p50 (macOS), so the waves are not runner-starved. The limit is the build depth.
• Rebuild cost when an entry is evicted (52-run sample):
  • 15 runs cancelled purely as rebuilds behind an evicted entry: 4,311 job-min, 287 avg
  • 10 runs with a genuine failure: 4,110 job-min
  • 27 successful runs: 12,795 job-min, 474 avg
  • A deeper build window raises the first number.

Root cause

max_entries_to_build: 5 caps the speculative pipeline at 5 entries. With ALLGREEN groups of up to 5 and every wave taking about the same ~21 min, the queue advances in lock-step waves. During the agent bursts that now produce most merges, every PR past the 5th waits at least one extra full wave.

Proposed fix (ruleset 24668460, merge_queue rule)

1. Raise max_entries_to_build 5 → 8, and keep max_entries_to_merge at 5 (or raise it to 8 too, so a fully green window lands in one merge).
2. Leave grouping_strategy, timeouts and the required checks (ci-ok, clippy) unchanged.
3. Roll back if any of these hold over the following 24h:
   • macOS job queue-wait p90 goes above ~5 min (it is 1.8 min now)
   • the merge_group eviction rate stays above ~20%. Then the extra rebuilds cost more than the latency buys; prefer landing the ci-janitor flake fixes (Fix Windows flake in read-set replacement test #1237, Retry HTTP errors in sbt/composer Docker image downloads #1238, Retry Docker e2e maven/cargo fixture fetches past registry blips #1222, Retry Bun backtest cells on hosted-fetch 5xx blips #1244) first.

Expected saving

• Latency: ~21 min off enqueue → merge for every PR that is 6th–8th in line. In the window above that was 8 of 13 PRs. Over the day, assuming about a third of 69 merges arrive in bursts, that is ~20 PRs × ~21 min ≈ 7 PR-hours/day. Merge-queue p50 should fall from ~64 toward ~45 min during bursts.
• Critical path per affected PR: −21 min (one wave).
• Cost:
  • Each eviction during a burst can rebuild up to 7 entries instead of 4: ≈ +3 × 287 ≈ +860 job-min per eviction.
  • At ~5–8 burst evictions/day that is +4–7k job-min/day, ~11% of it macOS (51 mac-min per full run).
  • Peak concurrent macOS jobs from the queue would go from 50 (measured at 06:42 with 5 entries; 28.9 mac jobs per run) to roughly 80.
  • CI perf: ci.yml e2e-macos — 23 one-minute macOS jobs per merge_group/push run (~2,300 macOS job starts/day) #1176 / Run e2e suites that share a toolchain in one leg #1234 (fewer e2e-macos legs) would offset most of that.

Coverage and risk

• No test moves; every job still runs on every merge_group run.
• Required check names are unchanged.
• Main risk: macOS pool pressure. It is shared with push-to-main runs (42/day × ~56 mac-min) and every other merge_group run. Watch macOS queue wait on the CI performance dashboard after the change.
• Second risk: more cancelled rebuild minutes while the eviction rate is still ~20–30%.

Effort

S (one ruleset field, reversible in seconds).

ROI

About 7: (≈21 critical-path min − ≈7 weighted for +5k job-min) × confidence 0.5 ÷ effort 1. Same scale as the dashboard.