You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository was archived by the owner on Nov 21, 2026. It is now read-only.
Repository navigation
This repository was archived by the owner on Nov 21, 2026. It is now read-only.
test_async_executor_bulk_command_is_exact_at_parallel_four stored 8742 of 9742 once on a darwin/arm64 runner #99
On 2026-10-02 the Test Python Bindings job for darwin/arm64, Python 3.11, on c4fdf9b failed with assert 8742 == 9742: one batch of 1,000 missing. The test is the regression test for ArcadeData#7615 (records dropped silently by the async executor at parallel level 4, fixed by ArcadeData#7625 for 26.10.1). The failure stopped at the count assertion, so the log does not say whether the batch was dropped with no callback (ArcadeData#7615 back) or rolled back and reported through on_error after its retries ran out, which ArcadeData#7625 allows under heavy contention.
Not reproduced locally on the current engine: 0 of 140 runs of the same load (all 16 cores, 3 cores, and 2 cores shared with two busy loops). The failed job was re-run. The fix here makes the count assertion print the on_error calls, so a recurrence says which of the two it is; if it is the first, it goes upstream with a Java reproduction.
Cause found: this is the upstream behaviour tracked in ArcadeData#9694. With the default retry settings a periodic batch commit on a one-bucket type at parallel level 4 can exhaust its retries and roll back the batch, about 0.4% of runs of this scenario even after upstream's ArcadeData#9535. Keeping this open until ArcadeData#9694 is fixed and verified; the test is right to fail on it.
ArcadeData#9694 is fixed by ArcadeData#9703 (merged into upstream main as 49b4fe2, in 26.11.1) and verified there: 0 lost batches in 13,000 runs of the Java reproduction across JDK 21, JDK 25, no per-command callback and four buckets (details on ArcadeData#9694). Keeping this open until the bindings pick up an engine build that contains it; then this test should stop failing on that runner and this issue closes.
Activity
tae898 commentedon Oct 10, 2026
Cause found: this is the upstream behaviour tracked in ArcadeData#9694. With the default retry settings a periodic batch commit on a one-bucket type at parallel level 4 can exhaust its retries and roll back the batch, about 0.4% of runs of this scenario even after upstream's ArcadeData#9535. Keeping this open until ArcadeData#9694 is fixed and verified; the test is right to fail on it.
tae898 commentedon Oct 11, 2026
ArcadeData#9694 is fixed by ArcadeData#9703 (merged into upstream main as 49b4fe2, in 26.11.1) and verified there: 0 lost batches in 13,000 runs of the Java reproduction across JDK 21, JDK 25, no per-command callback and four buckets (details on ArcadeData#9694). Keeping this open until the bindings pick up an engine build that contains it; then this test should stop failing on that runner and this issue closes.
docs: weekly audit fixes 2026-10-13 (#317)
docs: the fixes that shipped in 26.10.2 are no longer "not released y…