rkhachatryan commented on a change in pull request #13827:
URL: https://github.com/apache/flink/pull/13827#discussion_r524193666
##########
File path:
flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/io/SingleCheckpointBarrierHandler.java
##########
@@ -235,6 +257,23 @@ int getNumOpenChannels() {
return numOpenChannels;
}
+ private CheckpointBarrier maybeTimeout(CheckpointBarrier barrier) {
+ CheckpointOptions options = barrier.getCheckpointOptions();
+ boolean shouldTimeout = (options.isTimeoutable()) && (
+ barrier.getId() == timeoutedBarrierId ||
+ (System.currentTimeMillis() - barrier.getTimestamp()) >
options.getAlignmentTimeout());
Review comment:
We discussed offline two options of switching a checkpoint from aligned
to unaligned:
1. after some time since the start of a checkpoint on JM
1. after some time since the start of alignment on this subtask
(in either case, all the downstream tasks would proceed in UC mode with this
checkpoint after it's timed out)
The latter gives more control, i.e. the ability to address long alignment in
isolation, but not long barrier travel time (e.g. long sync phase or just long
pipeline).
We couldn't agree on which one would be more convenient for the users.
@NicoK, do you have an opinion on that?
Or maybe we should provide both?
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
