Hi devs,

I recently opened a pull request to fix a bug in HybridSourceReader
described in FLINK-30514. Since it has not received any review yet, I
would appreciate it if someone familiar with the connector-base module
could take a look.

The issue: During source switching, HybridSourceReader calls start()
before addSplits(), which breaks the assumption of underlying source
readers (e.g., IcebergSourceSplitReader) that recovered splits will be
available before start() is invoked. This causes undesired split
requests during recovery, breaking watermarks and endangering job
stability.

The fix is small (58 additions, 7 deletions) -- it reorders the method
calls so that start() is called after recovered splits are added. Unit
tests have been updated and CI is green.

JIRA: https://issues.apache.org/jira/browse/FLINK-30514
PR: https://github.com/apache/flink/pull/28069

Any feedback or review would be greatly appreciated. Thanks in advance!

Best,
Chase

Reply via email to