Repository navigation
perf(ci): cut superseded-head runs — 64% of runner-minutes on last 15 merged PRs #5430
Description
Measured 2026-10-16 (GraphQL check runs, last 15 merged PRs, last 8 commits each): 1398 runner-min on non-final heads vs 775 on final heads. Superseded: 35 CANCELLED jobs = 967 min (avg 28 min wasted before cancel), 68 FAILED jobs = 572 min. cancel-in-progress already on (ci.yml concurrency); loss is pushes mid-run. Countermeasures, no workflow edit: (1) push only after local green (tier2 + guard-tree locally) [existing rule]; (2) no main-merge/update-branch push onto a PR with CI running or seats pending unless BEHIND blocks arming; (3) batch fixups into one push; (4) pushes/PR counted in the hourly line (target <=2; today median 3, max 18). Owner: a3/sweep enforce. Refs aprender-pace PACE-8.
Activity
noahgift commentedon Oct 16, 2026
aprender-pace, 2026-10-16: cancel-in-progress is ALREADY on (ci.yml l.51-53, pull_request + merge_group), so it is not the lever: 48 merged PRs spent 5467 of 9816 runner-min (56%) on superseded heads, avg 28 min run before the cancel. The loss is the push that lands mid-run. One countermeasure: PUSH BATCHING per PR. (1) push only after local green of the same checks (one push per fix round); (2) no main-merge/update-branch while the head's CI is running — merge main only when the head is red/finished or CONFLICTING; (3) fixups batched into one push; (4) a3/sweep count pushes per PR and report >3 per PR on the hourly line. Target: median 3 -> 1.5 commits/PR halves the 56%. No workflow edit needed. Owner: a3 (enforce) + sweep.
noahgift commentedon Oct 22, 2026
Tracked in #5510 (triage 2026-10-22).