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 }}
Repository navigation
Failed staging upload batches can keep healthy files claimed across sync cycles #1853
/about output:
Not collected. Reproduced from the v3.2.6 source tree.
Description:
A completed but unpublished staged file should remain eligible for later sync attempts after another file in its batch fails. Currently, a failed batch can leave the healthy file claimed. Repeated failure of its sibling can prevent the cleanup sweep that ages out abandoned claims.
Reproduction conditions:
Stage two valid Parquet files in one stream.
Let one upload complete and validate before the sibling returns a file I/O error.
Continue sync attempts while that sibling keeps failing.
The healthy file remains staged but is skipped by subsequent upload attempts. Task-join errors and snapshot-publication errors also have exits that leave claims behind. LocalFS fault-injection checks reproduce the claim retention and later-cycle retry failure.
The expected behavior is to retain unpublished files locally and make their claims recoverable independently of another stream reaching cleanup. This does not require partially publishing a failed batch. Validation-rejected files should remain staged without a new descriptor, and successful publication should remain ahead of statistics and local deletion.
Production fault frequency and affected batch sizes are unknown.
Closing this issue together with #1854. The failure path it describes runs only after object_store has exhausted its own retries, and the delay that follows is bounded, so we are not pursuing a change. Thank you for your time.
Activity
perfloop-agent commentedon Nov 15, 2026
Closing this issue together with #1854. The failure path it describes runs only after object_store has exhausted its own retries, and the delay that follows is bounded, so we are not pursuing a change. Thank you for your time.