Martijn Visser created FLINK-40779:
--------------------------------------
Summary: JdbcSourceEnumerator snapshots splitter state ahead of
the splits it has handled
Key: FLINK-40779
URL: https://issues.apache.org/jira/browse/FLINK-40779
Project: Flink
Issue Type: Bug
Components: Connectors / JDBC
Affects Versions: jdbc-4.1.0
Reporter: Martijn Visser
{{JdbcSourceEnumerator}} runs {{SplitterEnumerator#enumerateSplits}} through
{{callAsync}}, so the splitter advances its state on the worker thread while
the splits only reach {{unassigned}} once the handler runs on the coordinator
thread. A checkpoint taken in between stores the new splitter state without
those splits, and they are lost on restore. {{SqlTemplateSplitEnumerator}} in
continuous mode is affected today.
{{snapshotState}} should use the splitter state that came back together with
the last handled splits, captured inside the callable, instead of calling
{{serializableState()}} directly.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
