Repository navigation
recordengine: TestMergeWritersHoldAdmittedShare saw writers report above their limit on Windows #748
Description
The failure
TestMergeWritersHoldAdmittedShare (recordengine/mergewriters_internal_test.go, added by #729) failed once on Windows CI in run 36474815871, at #747's commit 8599a77. #747 doesn't touch the record engine's merge code, and the next run passed.
mergewriters_internal_test.go:128: unbounded: writers report 93.4 MiB, heap 110.6 MiB
mergewriters_internal_test.go:129: share 32.0 MiB: grant 71.9 MiB, writers limited to 27.2 MiB, report 27.6 MiB (one append 10.4 MiB), heap 45.2 MiB
Error: "28907943" is not less than or equal to "28486562"
Why it may not be only noise
The failed assertion compares the writers' own reported resident bytes with their limit, not a GC-sensitive heap sample. The report was 27.6 MiB against a 27.2 MiB limit, about 0.4 MiB over. So in this run, the writers held more than their share, as they themselves report.
Candidate causes:
- A sequence that depends on timing, where one append (10.4 MiB here) is admitted against a stale reading.
- A platform difference in how the resident bytes are sampled.
Either way, #729's contract is that the limit is never exceeded. So this needs a look before it is written off as flaky.
To do
- Reproduce it with
-countand a stress loop, on Linux and Windows if possible. - Find which append crossed the limit, and whether the admission check read a stale value.
- Fix the bound, or document the exact slack the contract allows and assert against that.
Related: #729.
🤖 Generated with Claude Code
Activity