ldadima commented on code in PR #24784:
URL: https://github.com/apache/flink/pull/24784#discussion_r1600910971
##########
flink-runtime/src/main/java/org/apache/flink/runtime/checkpoint/StateAssignmentOperation.java:
##########
@@ -421,7 +423,27 @@ public void
reDistributeInputChannelStates(TaskStateAssignment stateAssignment)
stateAssignment.oldState.get(stateAssignment.inputOperatorID);
final List<List<InputChannelStateHandle>> inputOperatorState =
splitBySubtasks(inputState,
OperatorSubtaskState::getInputChannelState);
- if (inputState.getParallelism() ==
executionJobVertex.getParallelism()) {
+
+ boolean noNeedRescale =
+
stateAssignment.executionJobVertex.getJobVertex().getInputs().stream()
+ .map(JobEdge::getDownstreamSubtaskStateMapper)
+ .anyMatch(m ->
!m.equals(SubtaskStateMapper.FULL))
+ &&
stateAssignment.executionJobVertex.getInputs().stream()
+ .map(IntermediateResult::getProducer)
+ .map(vertexAssignments::get)
+ .anyMatch(
+ taskStateAssignment -> {
+ final int oldParallelism =
+ stateAssignment
+ .oldState
+
.get(stateAssignment.inputOperatorID)
+ .getParallelism();
+ return oldParallelism
+ ==
taskStateAssignment.executionJobVertex
+ .getParallelism();
+ });
Review Comment:
Separation is a good idea, ok.
The first part is to check if SubtaskStateMapper.FULL is used in any input
(this mapper type always returns all I/O), the problem is reproduced only for
this mapper type
The second part is to check if the parallelism of any previous operator has
changed. if it has changed, it means that the number of outputs has changed.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
