Skip to content

test-infra: transcription-do 'voice tier' flakes under full-suite vitest-pool-workers concurrency (double-metered minutes) #171

Description

@unforced

Found during C2 review (PR #170, round 2). ~30-40% of FULL-suite bun x vitest run runs in workers/vault fail the voice-tier test with double-metered minutes (596 vs 598). Never reproduces running the file in isolation.

Bisected properly: reproduced at the identical rate at the pre-round-2 commit (4d4926f) AND traced to a TOCTOU window in transcribeOne's "is this attachment still pending" check — only reachable when the vitest pool interleaves two alarm() invocations under load. Real Cloudflare DOs serialize alarm invocations, so no production exposure is believed — but the C2 reviewer is delivering an independent judgment on that claim (will be appended here).

Impact: can occasionally flake an unrelated PR's CI. Options: (a) serialize alarm invocations in the test harness to match real DO semantics, (b) make transcribeOne's pending-check compare-and-set atomic (also hardens production against future workerd semantic changes), (c) both.

Related: the alarmRunning reentrancy guard added in #170 narrows but does not close this window under the pool's concurrency model.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions