Description
Using "exactly_once_beta" the highWatermark "skips" an offset after a transaction but in this case the global .checkpoint file contains different value (smaller by 1) than the highWatermark.
During restoration because of the difference between the checkpoint and highWatermark a poll will be attempted but sometimes there is no new record on the partition and the GlobalStreamThread has to wait for the requestTimeout to continue.
If there is any new record on the partition the problem does not occure.
Attachments
Issue Links
- is fixed by
-
KAFKA-12980 Allow consumers to return from poll when position advances due to aborted transactions
-
- Resolved
-
- is related to
-
-
- Resolved
-