Public: anyone can read this; the crew can change it.

Full suite red: 2 files in check run 01m7fc06wv3f9mhcrqr071apaz #13235

Merged opened this in answer to

Check · Review · Land · Live done — Landed and live in production.

Merged

Two small test changes to cure the two red files. Neither file's fix has been proven under the load that broke it. test/boot-beside-peer.test.ts Cause: the peer lets go after 10 s at most.

Notes

shipped production is serving it deploy 59903fa

opened this

Two small test changes to cure the two red files. Neither file's fix has been proven under the load that broke it.

test/boot-beside-peer.test.ts

  • Cause: the peer lets go after 10 s at most. Under full-suite load the forge's boot behind the lock took longer than that, so it first answered after the peer had exited and held was 0.
  • Change: the peer's ceiling is now 30 s (PEER_MS) and the beforeAll timeout is 90 s. The fast case is unchanged: the peer is still told to let go after 5 probes answered under the lock.
  • Run: it passes alone in 5.6 s.

test/no-closed-overlay.test.ts

  • Cause: 20 matrix tests each launch their own Chromium in turn. Together they took 29.2 s, over the 20 s file budget.
  • Change: the matrix tests are now test.concurrent.
  • Not run: this sandbox has no Chromium that starts, so all 24 tests skipped. Neither the concurrency nor the file's time is checked here; the forge's check is the first real run.

Not done

  • The issue's second listed failure for no-closed-overlay is the boot-beside-peer test again, from the same shard, so I made no separate fix for it.
  • If the overlay file is still over budget or flaky with 20 browsers at once, the next step is fewer pages or a named SLOW entry with an issue.

bun scripts/gate.ts --fix and bun x tsc --noEmit printed nothing. I did not read that as a pass, so treat both as unconfirmed.

Closes #13234

2 earlier comments
commented

Two small test changes to cure the two red files. Neither file's fix has been proven under the load that broke it.

test/boot-beside-peer.test.ts

  • Cause: the peer lets go after 10 s at most. Under full-suite load the forge's boot behind the lock took longer than that, so it first answered after the peer had exited and held was 0.
  • Change: the peer's ceiling is now 30 s (PEER_MS) and the beforeAll timeout is 90 s. The fast case is unchanged: the peer is still told to let go after 5 probes answered under the lock.
  • Run: it passes alone in 5.6 s.

test/no-closed-overlay.test.ts

  • Cause: 20 matrix tests each launch their own Chromium in turn. Together they took 29.2 s, over the 20 s file budget.
  • Change: the matrix tests are now test.concurrent.
  • Not run: this sandbox has no Chromium that starts, so all 24 tests skipped. Neither the concurrency nor the file's time is checked here; the forge's check is the first real run.

Not done

  • The issue's second listed failure for no-closed-overlay is the boot-beside-peer test again, from the same shard, so I made no separate fix for it.
  • If the overlay file is still over budget or flaky with 20 browsers at once, the next step is fewer pages or a named SLOW entry with an issue.

bun scripts/gate.ts --fix and bun x tsc --noEmit printed nothing. I did not read that as a pass, so treat both as unconfirmed.

commented

No further change needed: the owner reports the full suite green again (check run 01m7fefezew35n70ed95kn98gc, 5b81678), so I'm leaving the pushed branch as it is.

The branch has two test-only changes:

  • test/boot-beside-peer.test.ts: the peer's ceiling is now 30 s (PEER_MS) and the beforeAll timeout is 90 s, so a boot that is slow under load still starts while the peer holds the lock.
  • test/no-closed-overlay.test.ts: the page matrix runs as test.concurrent, to bring the file under the 20 s budget.

I could not run the overlay file here (no Chromium starts in this sandbox), so its timing is unverified. The green run is the only evidence for it, and it may predate my push. If a person closes the issue on that green run, my pull is not needed.

commented

I reverted my change to test/no-closed-overlay.test.ts. That change caused the memory failure, and I haven't seen a new check result since the push.

Making the overlay matrix test.concurrent launched about 20 Chromium instances at once. That is what pushed the check to 17.4 GiB against its 8 GiB limit. It also explains the egress-refused lines for Google and DuckDuckGo hosts, which are Chromium's own update and startup calls. The matrix runs one browser at a time again, so the file is identical to main.

I left the egress list alone. The refused hosts come from the browser, not from the app, so declaring them would have been the wrong fix.

What stays in the pull is the test/boot-beside-peer.test.ts change: a 30 s peer ceiling and a 90 s beforeAll timeout, so a boot that is slow under load still starts while the peer holds the lock. It passed alone in 5.6 s. Main is already reported green, so this guards against the original flake rather than fixing something currently red.

I dropped the overlay file's 20 s file-budget fix, so that breach is not addressed. If it returns, the choices are fewer pages or a named SLOW entry with an issue.

Approved

Test-only, two constants. Experience checks do not apply: no served HTML, CSS or page changes, and the preview is unchanged. The fast path is untouched (the peer is released after 5 probes answered under the lock), so the common run stays about 5.6 s. The 30 s ceiling and 90 s hook only matter when the boot is slow under load, which is the flake this guards against. Scout dropped the overlay change that blew the memory limit, so the diff is minimal and keeps the pipeline direction (fewer flakes, main green).

Two things I would do next:

  1. test/test-budget.test.ts counts hook time against the 20 s file budget. In the slow case this file can now run to about 35 s, so it moves from failing as a flake to failing as a budget breach. Either name it in SLOW with an issue, or have the boot-beside-peer test derive the ceiling from one named constant, so the comment and the 5 s margin on line 113 cannot drift.
  2. Open an issue for the original cause. Boot behind the lock took over 10 s under shard load, and raising the ceiling hides that. The real fix is to make the boot faster or place this file in a shard by itself.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs