Details
-
Bug
-
Status: Closed
-
Minor
-
Resolution: Fixed
-
1.14.4, 1.15.0, 1.16.0
Description
On recovery, HeapRestoreOperation reads state handles one by one;
- each handle contains metadata at the beginning;
- the metadata is always read, but not actually used if a state with the corresponding name was already registered
In a rare case of downscaling + multiple checkpoints with different metadata; this might lead to data being deserialized incorrectly (always using the initial metadata).
It also prevents incremental checkpoints with schema evolution.
On first access, however, the backend itself will update (merge) metadata; so that it doesn't affect new state updates.
Attachments
Issue Links
- is duplicated by
-
FLINK-28553 The serializer in StateMap has not been updated when metaInfo of StateTable updated
- Closed
- is related to
-
FLINK-23143 Support state migration
- Closed
- relates to
-
FLINK-21648 FLIP-151: Incremental snapshots for heap-based state backend
- Open
- links to