Hong Liang Teoh created FLINK-36820:
---------------------------------------
Summary: SourceCoordinator#addSplitsBack() should not be called on
full failover
Key: FLINK-36820
URL: https://issues.apache.org/jira/browse/FLINK-36820
Project: Flink
Issue Type: Improvement
Components: Connectors / Common
Affects Versions: 2.0-preview, 1.20.0
Reporter: Hong Liang Teoh
The
[SourceCoordinator|https://github.com/apache/flink/blob/2897ab7db1081c9bf1f477dc221b84e7c6686a6d/flink-runtime/src/main/java/org/apache/flink/runtime/source/coordinator/SourceCoordinator.java#L376-L382]
calls addSplitsBack() on restart to return uncheckpointed splits back to the
coordinator upon job restart.
This makes sense in the "pipelined" restart strategy.
However, the
[SourceCoordinator|https://github.com/apache/flink/blob/2897ab7db1081c9bf1f477dc221b84e7c6686a6d/flink-runtime/src/main/java/org/apache/flink/runtime/source/coordinator/SourceCoordinator.java#L376-L382]
should not call addSplitsBack() if the job is configured on "full" failure
recovery instead of "pipelined", because the operator state would be reset
fully in that case anyways.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
