Repository navigation
ci(coverage): run the coverage job on a sixteen vCPU runner - #4664
Conversation
Size Report
Startup median (7 runs, lower is better):
|
|
The change at b004cb9 is ready: it only switches the Coverage job to the 16 vCPU runner label, and that job passed on it at this head (run 44175823688). All other checks except one pass, and that one is still queued, not failed. There are no conflicts. Before merge, #4662 needs to land first and the queued check needs to finish. Not blocking, take or leave: the three-line comment at ci.yml#L293 repeats the PR body's measurements, which will go stale, so one line like "16 vCPU: the suite is CPU-bound; halves wall time at about equal billed minutes (#4664)" would do; and one green run (146 s) does not show 16 vCPU variance against the 8 vCPU range of 214-389 s, so re-running Coverage two or three more times on this head would show whether a larger runner exposes tests that assume small pids or fixed /proc reads, as #4662 did for three host-kit files. Runner rates and quota limits for 16 vCPU are not visible from the repo, so the cost side is a maintainer call. |
The job's suite is CPU-bound: instrumented execution and transform, with Vitest deriving its worker count from the machine. On eight vCPU the seven workers left the rest of the run queueing. The same gate on a sixteen vCPU runner finished the run step in 146 s, against 298 s at the median of green runs on eight, for close to the same runner cost. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
b004cb9 to
02cc380
Compare
Apex Sentinel · ✅ Approve · review of
|
| Command | What it does |
|---|---|
@apex-sentinel-staging note <context> |
Adds context; Sentinel re-checks the cited code |
@apex-sentinel-staging pivot <why> |
Says the finding is wrong; Sentinel re-checks the cited code |
@apex-sentinel-staging like |
Marks the finding as useful |
@apex-sentinel-staging dislike <why> |
Marks it as not useful, so it is not raised again |
On this conversation, comment @apex-sentinel-staging review to run a full review once again.
|
Applied both notes.
Range 146–176 s for the step, 172–209 s for the job. The slowest sample is still under the 8 vCPU p90 of 360 s, and under its max of 389 s. All four green, with Not seen: the |
Summary
The
Coveragejob ran onblacksmith-8vcpu-ubuntu-2404. Its suite is CPU-bound — instrumented execution plus transform — and Vitest takes its worker count from the machine, so seven workers left the rest of the run queueing. Moving the job toblacksmith-16vcpu-ubuntu-2404doubles the workers.One
runs-online in.github/workflows/ci.yml, plus a comment. No gate, threshold or verdict changes: the sameunit-ci,coverage-model,coverage-crap-modelandtmpdir-leaksgates run in the same order, with the changed-line and CRAP steps still reusing the one lcov.Depends on #4662, which has to merge first. Three
host-kitfiles let the runner answer for a fixture pid through/proc/<pid>/stat; those pids were unused on an 8 vCPU runner and used on a 16 vCPU one, so the job would go red at exactly this head.Validation
Measured in CI at identical gates,
Run coveragestep and whole job from the Actions API.Run coverage16 vCPU samples: run 44175823688 job 1, then attempts 1–3 of 44186117469. The spread is 146–176 s against 214–389 s on 8 vCPU, so the worst 16 vCPU sample still beats the 8 vCPU p90. Every gate step was green in all four. 16 vCPU halves the wall time for about 10% more money; 32 vCPU was green only in the probe lane and costs 31% more, so 16 is the pick.
Head
02cc3803c4.node --experimental-strip-types scripts/gate/check.ts— manifest ok. Unresolved risk: thecoordinate.contract.test.ts5 s timeout seen once on #4660 was not reproduced in these four runs, on 8 or 16 vCPU.