Repository navigation
design(ci): small PRs queue the full matrix and starve lightweight workflows #2243
Description
Problem. Each PR starts about ten workflows. A batch of small fix PRs pushed the Actions queue past 100 waiting runs on 2026-10-26. Meanwhile:
- lightweight workflows waited for the account's runner slots behind 45-minute test legs:
milestone.yml(ci: set milestones automatically on close #2218), Project Status and CI Reports; - the milestone for a newly opened item lagged by more than an hour, so it was set by hand.
Options
- A. No workflow change: maintainers open PRs in small batches while the queue is short. This is the stopgap in use now.
- B. Docs-only and test-only PRs skip the heavy matrices, using path filters or the existing
classify_changesgate, so they run only the jobs their files can affect. - C. Cancel superseded PR runs on a new push (
concurrencywithcancel-in-progressper PR).
Recommendation: B, plus C for PR events only. B cuts the volume at the source. C stops amended PRs from queuing stale runs. Neither touches main's push runs or the required checks.
Activity
JarryShaw commentedon Oct 26, 2026
Ruling (owner, in chat): fix it properly with option B. Docs-only and test-only PRs skip the heavy matrices.
mainpush runs andRequired checks passed.Option C (cancelling superseded runs) is not part of this ruling.
JarryShaw commentedon Oct 26, 2026
Why option C is out (owner, in chat): a cancelled superseded run leaves a red X on its commit. That is confusing, and useless for navigating at a glance. It matches what
mainshows since #1905: commits whose only non-success checks are cancelled Experimental runs render as failed. So the fix is B alone: skip the heavy jobs, never cancel them.ci: let docs-only and test-only pull requests skip the heavy matrices (…