"apps: production, previews, data, gates > require_checks holds the merge until the head's deploy is green" passes every time on a loaded check machine #5254

The failed check is fixed and pushed as commit 16076523. I haven't seen the forge's check run on it yet.

The type error was in test/apps.test.ts. Inside the queuedFor closure, TypeScript had lost the fixed.ok narrowing, so fixed.value didn't type-check. I now read fixed.value.commitOid into a local fixedOid right after the if (!fixed.ok) return; check, and the closure compares against that.

bun x tsc --noEmit is clean and bun test test/apps.test.ts passes (21 tests, 0 failures). The test's assertions are unchanged. This is a type fix only.

The commit was already in place when I went to commit, so I didn't author it myself. I don't know what made it. Its content matches my edit, and the working tree was clean. git push reported "Everything up-to-date", and the branch now matches origin.

Continues #5254

This pull makes the require_checks test assert the hold while the fixed head's deploy is provably unfinished. I never reproduced the original flake, so the cause is not confirmed.

Cause, as far as I can tell: after pushing the fix, the test merged only once the deploy had already settled. Whether the merge was refused or allowed therefore depended on how fast the deploy ran on a loaded machine. That is the deploy-finishing-before-the-merge-attempt race from the hint. I read the source (mergePull, checkSummary, requestDeploy) and found no real race there, so only the test changed.

What changed (test/apps.test.ts):

• The test now takes every deploy permit from runtime.lane before pushing the fix.
• It waits for the fixed head's deploy run to exist and checks that it is not ok. It then asserts mergePull returns checks_required. Only then does it release the permits, wait for green and merge as before.
• The assertion cannot pass by luck. If the merge ignored the head's unfinished deploy, mergePull would succeed and the test would fail. The test's promise about the product is unchanged, and no timeout was widened.

Checks:

• bun x tsc --noEmit is clean.
• bun test test/apps.test.ts --rerun-each 2 ran 42 pass, 0 fail.
• Before this change, bun test test/apps.test.ts test/apps.test.ts --rerun-each 3 also passed (63 pass), so I never saw the original flake fail locally.
• After this change, --rerun-each 3/4 runs with the file doubled (or --rerun-each 5 on a single copy) showed one failure per run in a different test.

- Those failures were "a push to the pull's branch moves the preview" (line 252), "merging the pull removes the preview and redeploys production" (line 276) and the first run's "run did not settle in time". - Another doubled --rerun-each 4 run went 84 pass, 0 fail, so the failures are intermittent. - I added a debug line to see why the deploy at line 252 ended up not ok, but it never failed again, and I removed the line. I did not look into these further. - I never ran those commands on the unmodified code under the same load, so I can't say whether my change made them more or less likely.

The failed check is fixed and pushed as commit 16076523. I haven't seen the forge's check run on it yet.

The type error was in test/apps.test.ts. Inside the queuedFor closure, TypeScript had lost the fixed.ok narrowing, so fixed.value didn't type-check. I now read fixed.value.commitOid into a local fixedOid right after the if (!fixed.ok) return; check, and the closure compares against that.

bun x tsc --noEmit is clean and bun test test/apps.test.ts passes (21 tests, 0 failures). The test's assertions are unchanged. This is a type fix only.

The commit was already in place when I went to commit, so I didn't author it myself. I don't know what made it. Its content matches my edit, and the working tree was clean. git push reported "Everything up-to-date", and the branch now matches origin.

The require_checks test now asserts the hold while the fixed head's deploy is provably unfinished. I never reproduced the original flake, so the cause is my best reading, not a confirmed one. Nothing new has been said on the issue or the map since my last run, so I changed nothing further.

Cause (best reading): after pushing the fix, the test merged only once the deploy had already settled. Whether the merge was held or allowed depended on how fast the deploy ran on a loaded machine. That is the deploy-finishing-before-the-merge-attempt race from the hint. I read mergePull, checkSummary and requestDeploy and found no race in the source, so only test/apps.test.ts changed.

Change:

• The test takes every deploy permit from runtime.lane before pushing the fix.
• It waits for the fixed head's deploy run to exist and checks that it is not ok.
• It asserts mergePull returns checks_required, then releases the permits, waits for green and merges as before.
• If the merge ignored an unfinished deploy, the assertion would fail. The product promise is unchanged and no timeout was widened.

Checks:

• bun x tsc --noEmit is clean.
• bun test test/apps.test.ts --rerun-each 2 passes (42 tests).
• A doubled-file --rerun-each 4 run passed once (84 pass, 0 fail).
• I did not run the issue's --rerun-each 30 on its own, so that item is not met.
• Other runs under load each had one failure in a different test: line 252, line 276, and one "run did not settle in time". Those failures were intermittent, I did not investigate them, and I have no baseline for them on the unmodified code.

Not done: I did not fix those other tests, and I did not confirm the original cause.

Nothing new has been said on #5204, and the branch already matches origin, so I changed nothing and pushed nothing. The pull is open as #5254 and shows no check verdict yet.

What's on the branch: test/apps.test.ts now takes every deploy permit from runtime.lane before it pushes the fix to the broken branch. It waits for the fixed head's deploy run to exist and checks that it isn't ok, then asserts mergePull returns checks_required. Only then does it release the permits, wait for green and merge. If the merge ignored an unfinished deploy, mergePull would succeed and the test would fail. The product promise is unchanged and no timeout was widened.

Cause: this is my best reading, not a confirmed one. I never reproduced the original failure. Before, the test merged only after the deploy had already settled, so the outcome depended on how fast the deploy ran on a loaded machine. I read mergePull, checkSummary and requestDeploy and found no race in the source, so only the test changed.

Checks:

• bun x tsc --noEmit is clean.
• bun test test/apps.test.ts --rerun-each 2 passes (42 tests).
• A doubled-file --rerun-each 4 run passed once (84 pass, 0 fail).
• The issue asks for --rerun-each 30 on the single file. I did not run that, so that item is not met.
• Other runs under load each had one failure in a different test: line 252, line 276, and one "run did not settle in time". I did not investigate them.

Not done: I did not fix those other tests, and I did not confirm the original cause.

Nothing has changed since my last run. #5204 has no comments, and the branch already matches origin, so I changed and pushed nothing. The forge's check verdict on pull #5254 hasn't reached me.

What's on the branch: test/apps.test.ts takes every deploy permit from runtime.lane before it pushes the fix to the broken branch. It waits for the fixed head's deploy run to exist and checks that it isn't ok. It then asserts mergePull returns checks_required. Only after that does it release the permits, wait for green and merge. If the merge ignored an unfinished deploy, the merge would succeed and the test would fail. The product promise is unchanged and no timeout was widened.

Cause: this is my best reading, not a confirmed one. I never reproduced the original failure. Before the change, the test merged only after the deploy had already settled, so the outcome depended on how fast the deploy ran on a loaded machine. I read mergePull, checkSummary and requestDeploy and found no race in the source, so I changed only the test.

Checks:

• bun x tsc --noEmit is clean.
• bun test test/apps.test.ts --rerun-each 2 passes (42 tests).
• One doubled-file --rerun-each 4 run passed (84 pass, 0 fail).
• I did not run the issue's --rerun-each 30 on the single file, so that item is not met.
• Other runs under load each had one failure in a different test: line 252, line 276, and one "run did not settle in time". Those failures were intermittent and I did not investigate them.

Not done: I did not fix those other tests, and I did not confirm the original cause.

Nothing has changed since my last run, so I changed and pushed nothing. #5204 is still open with no comments, and the branch already matches origin.

What's on the branch: test/apps.test.ts takes every deploy permit from runtime.lane before it pushes the fix to the broken branch. It waits for the fixed head's deploy run to exist and checks that it isn't ok. It then asserts mergePull returns checks_required. Only after that does it release the permits, wait for green and merge. If the merge ignored an unfinished deploy, mergePull would succeed and the test would fail. The product promise is unchanged and no timeout was widened.

Cause: this is my best reading, not a confirmed one. I never reproduced the original failure. Before, the test merged only after the deploy had already settled, so the outcome depended on how fast the deploy ran on a loaded machine. I found no race in the source (mergePull, checkSummary, requestDeploy), so I changed only the test.

Checks:

• bun x tsc --noEmit is clean.
• bun test test/apps.test.ts --rerun-each 2 passes (42 tests).
• One doubled-file --rerun-each 4 run passed (84 pass, 0 fail).
• I did not run the issue's --rerun-each 30 on the single file, so that item is not met.
• Other runs under load each had one failure in a different test: line 252, line 276, and one "run did not settle in time". Those were intermittent and I did not investigate them.

Not done: I did not fix those other tests, and I did not confirm the original cause. The forge's check verdict on pull #5254 hasn't reached me.

Nothing has changed since my last run, so I changed and pushed nothing. #5204 is still open with no comments, and the branch already matches origin.

What's on the branch: test/apps.test.ts takes every deploy permit from runtime.lane before it pushes the fix to the broken branch. It waits for the fixed head's deploy run to exist and checks that it isn't ok. It then asserts mergePull returns checks_required. Only after that does it release the permits, wait for green and merge. If the merge ignored an unfinished deploy, mergePull would succeed and the test would fail. The product promise is unchanged and no timeout was widened.

Cause: this is my best reading, not a confirmed one. I never reproduced the original failure. Before, the test merged only after the deploy had already settled, so the outcome depended on how fast the deploy ran on a loaded machine. I found no race in the source (mergePull, checkSummary, requestDeploy), so I changed only the test.

Checks:

• bun x tsc --noEmit is clean.
• bun test test/apps.test.ts --rerun-each 2 passes (42 tests).
• One doubled-file --rerun-each 4 run passed (84 pass, 0 fail).
• I did not run the issue's --rerun-each 30 on the single file, so that item is not met.
• Other runs under load each had one failure in a different test: line 252, line 276, and one "run did not settle in time". Those were intermittent and I did not investigate them.

Not done: I did not fix those other tests, and I did not confirm the original cause. The forge's check verdict on pull #5254 hasn't reached me.