Skip to content

CI: one runner wait for the PR upgrade check, model lock on the 16-core pool - #1921

Merged
davidmckayv merged 1 commit into
mainfrom
ci/upgrade-signing-fast
Nov 20, 2026
Merged

davidmckayv merged 1 commit into
mainfrom
ci/upgrade-signing-fast

Conversation

@davidmckayv

Copy link
Copy Markdown
Contributor

What

Faster pull request runs for Desktop upgrade, Model lock and Desktop macOS signing. All three were mostly waiting for runners, not working.

Workflow (on a PR) Wall time before Work in it Where the time went
Desktop upgrade 17 to 22 min about 3.5 min two runner waits in a row: 3 to 6 min for the legs job, then 9 to 16 min for the upgrade job
Model lock (lock change, every harness) 32 min at most 7.5 min per harness 13 min for the first job, 8 to 12 min for each harness job, all on the standard pool
Desktop Windows signing dry run 0.5 to 8 min about 20 s already fast, left alone
Desktop macOS signing dry run 1 to 4 min about 1.5 min already fast
security / zizmor about 15 s already fast, left alone

Changes

  • Desktop upgrade: on a pull request the upgrade job finds the previous release itself (the same upgrade-legs.ts previous call), so the run waits for one runner instead of two in a row. The legs job runs only for the nightly and on-demand window, now on the 16-core pool. Nothing else about the test changes; with no earlier release the job does nothing, as before.
  • Model lock: both jobs move from the standard pool, which the organisation caps at 60 concurrent jobs, to the 16-core pool. A lock change starts a job per harness, and those queued behind every other workflow. Fork pull requests are still skipped.
  • Desktop macOS signing: a pull request's newer push cancels its superseded dry run so it stops holding one of the few macOS runners. Pull request runs never reach the signing job. Tag, main and dispatch runs, the ones that sign, still never cancel.
  • Desktop Windows signing: unchanged. A pull request labelled for signing does real signing on a pull_request event, so cancelling its runs would change signing behaviour.

Checks

  • actionlint (only the known custom runner label notice) and zizmor: no findings.
  • The workflow-reading tests pass: signing-workflows, signed-build, staple-with-retry, desktop-updates, release-latest, release-publish-order, model-compat-langgraph-ts.
  • This PR's own run exercises Desktop upgrade, Model lock (all harnesses, since the workflow changed) and Desktop macOS signing.

…re pool

Desktop upgrade on a pull request waited for two runners in a row: a
small job on the shared standard pool to name the previous release, then
the upgrade job. Both queued (3 to 6 and 9 to 16 minutes) around about
3.5 minutes of work. The upgrade job now finds the previous release
itself, and the separate legs job runs only for the nightly window.

Model lock started one job per harness on the standard pool, which the
organisation caps at 60 concurrent jobs; a lock change queued 8 to 13
minutes per job and took 32 minutes for at most 7.5 of work. Both of its
jobs now run on the 16-core pool.

Desktop macOS signing cancels a pull request's superseded dry run, so a
stale run does not hold one of the few macOS runners. Tag, main and
dispatch runs, the ones that sign, still never cancel.
@davidmckayv
davidmckayv merged commit 86c96df into main Nov 20, 2026
7 checks passed
@davidmckayv
davidmckayv deleted the ci/upgrade-signing-fast branch November 20, 2026 05:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant