[
https://issues.apache.org/jira/browse/FLINK-10966?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16695584#comment-16695584
]
zhijiang commented on FLINK-10966:
----------------------------------
I have some concerns about this issue:
# In credit-based flow control mode from release-1.5, the blocked channel will
cache the data in memory instead of spill to disk before. So the blocking
effect seems limited.
# If checkpoint id1 is timeout, checkpoint id2 is then trigged by coordinator.
The tasks would release previous blocked channels when receiving one new
checkpoint id, so as long as the back pressure is not very extreme, the new
checkpoint id should be arrived not very delay.
# The proposed PRC CancelCheckpointMarker message before triggering new
checkpoint id from coordinator can indeed speed to release blocked channels of
tasks in some scenarios, such as back pressure. But I just wonder is it
necessary to trigger another checkpoint before previous one success? In other
words, if the current checkpoint timeout, maybe the new triggered checkpoint
also timeout as a result. It has not solved the essential reason of timeout.
> Optimize the release blocking logic in BarrierBuffer
> ----------------------------------------------------
>
> Key: FLINK-10966
> URL: https://issues.apache.org/jira/browse/FLINK-10966
> Project: Flink
> Issue Type: Improvement
> Components: State Backends, Checkpointing
> Reporter: vinoyang
> Assignee: vinoyang
> Priority: Major
>
> Issue:
> Currently, mixing CancelCheckpointMarker control events with data flow to
> drive task to release blocking logic in BarrierBuffer may result in blocking
> logic not being released in time, further leading to a large amount of data
> being spilled to disk.
> The source code for this problem is as follows:
> {code:java}
> BufferOrEvent bufferOrEvent = next.get();
> if (isBlocked(bufferOrEvent.getChannelIndex())) { //issue line
> // if the channel is blocked we, we just store the BufferOrEvent
> bufferBlocker.add(bufferOrEvent);
> checkSizeLimit();
> }
> else if (bufferOrEvent.isBuffer()) {
> return bufferOrEvent;
> }
> else if (bufferOrEvent.getEvent().getClass() == CheckpointBarrier.class) {
> if (!endOfStream) {
> // process barriers only if there is a chance of the checkpoint
> completing
> processBarrier((CheckpointBarrier) bufferOrEvent.getEvent(),
> bufferOrEvent.getChannelIndex());
> }
> }
> else if (bufferOrEvent.getEvent().getClass() == CancelCheckpointMarker.class)
> {
> processCancellationBarrier((CancelCheckpointMarker)
> bufferOrEvent.getEvent());
> }
> else {
> if (bufferOrEvent.getEvent().getClass() == EndOfPartitionEvent.class) {
> processEndOfPartition();
> }
> return bufferOrEvent;
> }
> {code}
> Scenarios:
> Considering a simple DAG:source->map (network shuffle), the degree of
> parallelism is 10. The checkpoint semantics is exactly once.
> The first checkpoint: barriers of 9 source subtask are received by all map
> subtask. One of the source subtasks is blocked, resulting in the failure to
> send barrier. Eventually, the checkpoint will fail due to timeout. At this
> point, 9 corresponding input channel are blocked because they have received
> barrier.
> Second checkpoint: At this point, the special source subtask is still blocked
> and cannot send any events to downstream, while the nine input channels are
> still blocked. From the current implementation, the data or events it
> receives will not be processed, but will be stored directly. Therefore, the
> barrier of the downstream task will not be released. The only hope is that
> the cached data reaches the maximum limit.
> I think the main problem here is that we should not store data which comes
> from blocked input channels directly. Especially when one input channel is
> blocked by upstream and nine input channels are marked as blocked, we may not
> always be able to release the blocking.
> A better mechanism might be that we send notifyCheckpointFailed callback via
> CheckpointCoordinator, allowing each task to unblock itself. This mechanism
> can make the release of the old checkpoint align independent of the trigger
> of the new checkpoint. If the interval of the checkpoints are very long but
> the timeout is very short, then the effect of the optimization will be more
> obvious.
> Ultimately, we want to reduce unnecessary blocking and data spill to disk.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
