Github user tillrohrmann commented on a diff in the pull request:
https://github.com/apache/flink/pull/2629#discussion_r86340921
--- Diff:
flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/tasks/StreamTask.java
---
@@ -540,15 +540,12 @@ private boolean performCheckpoint(CheckpointMetaData
checkpointMetaData) throws
synchronized (lock) {
if (isRunning) {
+ checkpointState(checkpointMetaData);
- // Since both state checkpointing and
downstream barrier emission occurs in this
- // lock scope, they are an atomic operation
regardless of the order in which they occur.
- // Given this, we immediately emit the
checkpoint barriers, so the downstream operators
- // can start their checkpoint work as soon as
possible
+ // broadcast barriers after snapshot operators'
states.
operatorChain.broadcastCheckpointBarrier(
-
checkpointMetaData.getCheckpointId(), checkpointMetaData.getTimestamp());
-
- checkpointState(checkpointMetaData);
+
checkpointMetaData.getCheckpointId(), checkpointMetaData.getTimestamp()
+ );
--- End diff --
Changing the order of broadcasting the checkpoint barrier to downstream
operators and performing the actual checkpoint is very problematic. First
drawing the checkpoint and then sending the checkpoint barrier downstream will
add the latency of creating the checkpoint to the barrier for each operator.
This will considerably influence the overall checkpointing time, because the
individual checkpoints are effectively executed sequentially. Thus, I fear that
this change is not possible.
Is the problem that the `AsyncWaitOperator#Emitter` can emit results while
we're drawing the checkpoints? I think the proper way to solve this problem is
to let the output operation happen under the checkpoint lock. That way, you are
sure that no element will be outputted while the checkpointing is active.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
