[FLINK-24347] KafkaSource cannot checkpoint if the parallelism is higher than the partition number

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 1.14.0, 1.13.2
• Fix Version/s: 1.14.0, 1.13.3
• Component/s: Connectors / Kafka
• Labels: pull-request-available
• Assignee: Fabian Paul
• Reporter: Fabian Paul
• Created: 21/Sep/21 11:17
• Updated: 21/Sep/21 23:15
• Resolved: 21/Sep/21 23:14

Description

The KafkaSourceEnumerator signals all the readers if there are no more splits available and the readers shut down. In case the parallelism is higher than the partitions of the consumed topic there are not enough partitions to distribute and reader subtasks go immediately into FINISHED state.
Currently, it is not possible to checkpoint if parts of the job are finished. (this is lifted once FLIP-147 is by default enabled Flink 1.15+)

We should only signal to the readers to go into finished if the job is in bounded execution mode and keep them idling otherwise.

Issue Links

• links to: GitHub Pull Request #17327
• links to: GitHub Pull Request #17329
• links to: GitHub Pull Request #17330