Details
-
Bug
-
Status: Resolved
-
P2
-
Resolution: Fixed
-
None
-
None
Description
The FlinkRunner does not persist its view of the per-partition Watermark holds. This can lead to elements to be considered late after restoring from a savepoint or resuming a failed pipeline.
Similar to the approach in BEAM-6650, we can recover the Watermarks by iterating through the keys of the state backend during recovery.
Attachments
Issue Links
- is cloned by
-
BEAM-7188 LTS backport: FlinkRunner does not checkpoint partition view of watermark holds
- Resolved
- links to