Status
Discussion thread:
https://lists.apache.org/thread/4w36oof8dh28b9f593sgtk21o8qh8qx4
https://lists.apache.org/thread/t0bdkx1161rlbnsf06x0kswb05mch164
Vote thread: https://lists.apache.org/thread/0bh530j5ob11lzj48vpm883sqwgmstp8
JIRA: https://issues.apache.org/jira/browse/FLINK-29844
Released: 1.17
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
While restoring from a snapshot, StateBackend will resolve the schema compatibilty to decide where to go.
Currently, if users use a customized serializer to process data, they have to make sure that the schema compatibility in the old serializer (maybe in Flink library) meets the need as users want. Or else they have to modify TypeSerializerSnapshot#resolveSchemaCompatibility(TypeSerializer<T> newSerializer) of the old serializer. There are no ways for users to specify the compatiblity with the old serializer in the new serializer. It also makes scheme evolution not supported in some scenarios.
For exmaple, if users want to implement a customized serializer which is compatible with ValueSerializer, user have to modify the compatibility logic in ValueSerializer.ValueSerializerSnapshot#resolveSchemaCompatibility(TypeSerializer<T> newSerializer). (ValueSerializer is final so that it also could not be extended).
So reversing the direction of resolving schema compatibility could improve the usability of schema evolution.
Public Interfaces
Add an extra method in TypeSerializerSnapshot.java and provide a default implementation before removing the deprecated TypeSerializerSnapshot#resolveSchemaCompatibility(TypeSerializer<T> newSerializer).
@PublicEvolving public interface TypeSerializerSnapshot<T> { // Check whether the serializer is compatible with the old one. default TypeSerializerSchemaCompatibility<T> resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot) { // Return INCOMPATIBLE as default before removing the deprecated method. return TypeSerializerSchemaCompatibility.incompatible(); } }
Mark the old method as deprecated and provide a default implementation.
@PublicEvolving public interface TypeSerializerSnapshot<T> { @Deprecated default TypeSerializerSchemaCompatibility<T> resolveSchemaCompatibility(TypeSerializer<T> newSerializer) { // Use new method to resolve before removing this deprecated method. return newSerializer.snapshotConfiguration().resolveSerializerSchemaCompatibility(this); } }
Make the method abstract after removing the deprecated TypeSerializerSnapshot#resolveSchemaCompatibility(TypeSerializer<T> newSerializer).
@PublicEvolving public interface TypeSerializerSnapshot<T> { // Check whether the serializer is compatible with the old one. public abstract TypeSerializerSchemaCompatibility<T> resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot); }
Proposed Changes
Because TypeSerializerSnapshot is a public interface, we have several steps to migrate the logic.
Step 1
@PublicEvolving public interface TypeSerializerSnapshot<T> { // Check whether the serializer is compatible with the old one. default TypeSerializerSchemaCompatibility<T> resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot) { // Return INCOMPATIBLE as default before removing the deprecated method. return TypeSerializerSchemaCompatibility.incompatible(); } @Deprecated default TypeSerializerSchemaCompatibility<T> resolveSchemaCompatibility(TypeSerializer<T> newSerializer) { // Use new method to resolve before removing this deprecated method. return newSerializer.snapshotConfiguration().resolveSerializerSchemaCompatibility(this); } }
- Add an extra method (TypeserializeSnapshotr#resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot)) in TypeSerializerSnapshot.java as above, and return INCOMPATIBLE as default.
- Mark the original method as deprecated and it will use new method to resolve as default.
- Implement the new method for all built-in TypeserializerSnapshots.
For the new customized serializer, users could implement their own logic of resolving compatibility in their new serializer and it will be used by flink.
For the old customized serializer, if users haven't implement TypeSerializerSnapshot#resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot), all behaviors are still same as before.
if they implement it and remove their implementation of old method, the new logic will be used by flink as well.
Step 2
After several stable version, we need to:
- Remove TypeSerializerSnapshot#resolveSchemaCompatibility(TypeSerializer<T> newSerializer) and related implementation.
- Make all places where use TypeSerializerSnapshot#resolveSchemaCompatibility(TypeSerializer<T> newSerializer) to check the compatibility call Typeserializer#resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot).
- Remove the default implementation of the new method.
@PublicEvolving public interface TypeSerializerSnapshot<T> { // Check whether the serializer is compatible with the old one. TypeSerializerSchemaCompatibility<T> resolveSchemaCompatibility(TypeSerializerSnapshot<T> oldSerializerSnapshot); }
Compatibility, Deprecation, and Migration Plan
The deprecation and migration plan are listed above.
Test Plan
- Existing UT of resolving compatibility could verify whether the old behavior could work as before.
- Add extra UT to verify the correctness of customized serializer.
- Add ITCase and e2e test case to mock the user behavior.
Rejected Alternatives
Introduce a new method in TypeSerializer
Compred to introducing a new method in TypeSerializer, we could get benefits if we introduce it in TypeSerializerSnapshot:
- TypeSerializerSnapshot still owns the responsibility of resolving schema compatibility, TypeSerializer could just pay attention to its serialization and deserialization as before.
- It's very convenient to implement it based on current implementation by all information in TypeSerializerSnapshot and tools which is also helpful for users to migrate their external serializer.
Introduce a method without default implementation
Why not:
We should make sure that user jobs still work without modifying any codes before removing the deprecated method.
So we need to have two steps to complete the migration and make sure that the first version will not break change.