The GitHub Actions job "Flink CDC CI" on 
flink-cdc.git/fix-readded-table-stale-schema has failed.
Run started by GitHub user sd4324530 (triggered by sd4324530).

Head commit for run:
80ced256382366a557f9ccc63a84f7851601aa70 / Pei Yu <[email protected]>
[FLINK-40795][cdc-runtime] Fix re-added table with drifted schema not being 
re-created downstream

When a table is removed from the pipeline and later added back while its
upstream schema drifted (e.g. a column was dropped while out of capture),
the source re-snapshots it and emits a CreateTableEvent carrying the
latest schema. The regular SchemaCoordinator used to drop that event as
"redundant" whenever any schema was recorded in its state, because
SchemaManager never removes schemas of tables that left the pipeline.
As a result the downstream table was not re-created, and the stale
evolved schema kept being pushed back to SchemaOperators and sink
writers, coercing records into the stale shape and causing metadata
mismatch failures at the sink.

- Treat a CreateTableEvent as redundant only when the recorded schema
  is identical to the incoming one, so drifted re-snapshots are applied
  while snapshot-stage duplicates are still skipped.
- When a CreateTableEvent meets an existing evolved schema, re-create
  the downstream table if absent and align it with schema difference
  events, deducing the target schema per behavior: the incoming schema
  for EVOLVE / TRY_EVOLVE / EXCEPTION, the least common schema for
  LENIENT, and the frozen current schema for IGNORE. Only the resulting
  CreateTableEvent is propagated downstream.
- Guard evolved-schema state updates with the same redundancy check so
  alignment events are not applied to the state twice.

Generated-by: Claude Code 2.1.278

Report URL: https://github.com/apache/flink-cdc/actions/runs/36030572345

With regards,
GitHub Actions via GitBox

Reply via email to