Skip to content

Windows: the torn-row witness can miss its hold when the poll count's rename collides with the hold's read (intermittent) #138

Open

Description

@KyleAMathews

Windows e2e shard 9/10 failed once on #99 (CI run 43947367739, job 133580471927, head a07e6e9); it passed on the previous run (43941508359, e54f769) and in the scratch run 43805903246.

  • Unit: fixed seed 20261105: a capture row read while the worker wrote it
  • Failure: the torn-row witness was not reached: 0 hold(s) completed, 0 timed out, 1 batch(es) began (batch-2704-1.begun, batch-2704-1.polls, batch-2704-1.polls.2704.tmp)
  • What the files suggest (not confirmed): the leftover batch-2704-1.polls.2704.tmp is the watchdog's writeWhole temp file, so a rename onto .polls failed. On win32 a rename onto a file that another thread or process holds open fails (EPERM or EBUSY). The copy's hold (appendClosingRow in capture-setup.mjs) reads .polls every 10 ms while it waits. Neither the hold nor the watchdog's count wrote .reached or .unreached.
  • Hook: ORACLE_GAPS_TEST_TORN_ROWS (gap CLI: a run under load can end non-zero (an internal error) where it should complete #83). fix(oracle-gaps): no wall-clock cut decides an outcome #99 does not change its code; whether fix(oracle-gaps): no wall-clock cut decides an outcome #99's timing makes the collision likelier is not measured.
  • Rule: "Determinism under load covers witnesses" (C): a hold that is released by a count must not fail by a file race.

This issue was written by Isaac.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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