The GitHub Actions job "Flink Connector JDBC CI" on flink-connector-jdbc.git/FLINK-40779-splitter-state-checkpoint-race has failed. Run started by GitHub user Admaing (triggered by Admaing).
Head commit for run: 4cb6d9f12f594f39d6caa88733210d8bee59d9e5 / yongfu.gao <[email protected]> [FLINK-40779][connector-jdbc] Snapshot splitter state together with its splits JdbcSourceEnumerator enumerates splits through callAsync, so the splitter cursor advances on a worker thread while the produced splits only reach `unassigned` in the handler on the coordinator thread. Reading the splitter state directly in snapshotState() could therefore record a state that had already moved past splits that were still in flight; those splits were never re-enumerated after a restore. Capture the splits and the splitter state together inside the callable, and commit the state from the handler. Enumerations are limited to one in flight, so the committed state and the buffered splits stay in sync no matter how callAsync schedules the callables; relying on their completion order would not be safe. snapshotState() now uses that committed state. For the setSql path the state is also seeded from the provider when no state is configured. Stateful providers such as JdbcSlideTimingParameterProvider are mutated in place while enumerating, so a first-batch checkpoint taken with a null state could not roll the provider back on restore and the in-flight first batch was still lost. The fix sits in the enumerator rather than in the individual SplitterEnumerator implementations so that every implementation is covered (SlideTimingSplitterEnumerator is exposed the same way) and the PublicEvolving interface stays unchanged. On the error path the committed state lags behind, so a restore may re-enumerate splits that were already produced. That is at-least-once rather than data loss. Generated-by: DeepSeek Harness (deepseek-v4.1-flash) Report URL: https://github.com/apache/flink-connector-jdbc/actions/runs/38029587979 With regards, GitHub Actions via GitBox
