Backlog metrics do not showing up in FlinkRunner

[talatuyarer]

We have a Flink Job which does not emit backlog metrics. Actually metrics are emitting in KafkaIO. However I could not see them on Flink Metric system. Looks like Beam -> Flink wiring is broken. I set metricContext in Checkpointing phase which is the place metrics emit on UnBoundedReader.

@mxm @tweise @angoenka Could you please review my MR ? I tested in our env. It is working as expected.

[github-actions]

Checks are failing. Will not request review until checks are succeeding. If you'd like to override that behavior, comment assign set of reviewers

[talatuyarer]

assign set of reviewers

[apilloud]

assign set of reviewers

[github-actions]

No reviewers could be found from any of the labels on the PR or in the fallback reviewers list. Check the config file to make sure reviewers are configured

[apilloud]

Run Java PreCommit

[je-ik]

Could we add test to verify that the change brings the expected behavior?

The test might probably just verify that the metrics are available after running a simple pipeline with some source that updates the metrics.

[je-ik]

Could we add test to verify that the change brings the expected behavior?

The test might probably just verify that the metrics are available after running a simple pipeline with some source that updates the metrics.

There is an analogous test in SourceInputFormatTest.testAccumulatorRegistrationOnOperatorClose that can serve as an inspiration.

[je-ik]

This can be moved to a field and reused between calls to run and snapshotState.

[je-ik]

We have a Flink Job which does not emit backlog metrics. Actually metrics are emitting in KafkaIO. However I could not see them on Flink Metric system. Looks like Beam -> Flink wiring is broken. I set metricContext in Checkpointing phase which is the place metrics emit on UnBoundedReader.

@mxm @tweise @angoenka Could you please review my MR ? I tested in our env. It is working as expected.

I wonder, why do we need to write metrics into the context on checkpoint? The current code does so in the call to advance(), is that not sufficient?

[talatuyarer]

Hey @je-ik

I wonder, why do we need to write metrics into the context on checkpoint? The current code does so in the call to advance(), is that not sufficient?

Current code does not emit backlog metrics on advance. I see it is on checkpointing phase https://github.com/apache/beam/blob/master/sdks/java/io/kafka/src/main/java/org/apache/beam/sdk/io/kafka/KafkaUnboundedReader.java#L255

Do I miss anything ?

[je-ik]

If I understand that correctly (and I'm definitely not sure about that :)), then you refer to the time when metrics are compute in the source (KafkaSource in this case). But the wrapper should update the metric container on each call to advance, see [1]. I wonder why it is needed to do the same when doing checkpoint (though it sort might make more sense than to do that on each call to advance()).- I'm not saying it is wrong, I'd just like to understand the reason.

[1] https://github.com/apache/beam/pull/25554/files#diff-3f0d15eaf4f2c4979c6ac3fb42cd02819b1e5159ad7b28f7ac052ecd0bb21768L60

[talatuyarer]

If I understand that correctly (and I'm definitely not sure about that :)), then you refer to the time when metrics are compute in the source (KafkaSource in this case). But the wrapper should update the metric container on each call to advance, see [1]. I wonder why it is needed to do the same when doing checkpoint (though it sort might make more sense than to do that on each call to advance()).- I'm not saying it is wrong, I'd just like to understand the reason.

Backlog is reported from getCheckpointMark(), which is done by some other thread. Not sure why it is done there. But this is the main issue. Advance function runs on main thread and it has Metric context so I am able to see element count metrics. However Checkpoint thread does not have context thats why It can not emit metrics.

If I move reportBacklog() function in advance function i am able to see backlog metrics too. We could do that in advance(), but that would unnecessary overhead for every single record. :)

[je-ik]

No, I think there is good reason the backlog is computed on checkpoint only. :)
Yes, there are two threads, one thread runs the source and the other does the checkpoint. There is a lock (getCheckpointLock()) that ensures that this is consistent. If the source updates its internal metrics from the checkpoint thread, I would feel these metrics should be reported to flink on the first call to advance() - which is apparently not the case, I'm just a little struggling to see why. This might indicate some other bug somewhere.

[je-ik]

retest this please

[talatuyarer]

@je-ik Looks like this change is not working. Based on initial experiment. Whenever I enable beammetrics Pipeline stop processing. Do you have any suggestion ? I believe there is a concurrency issue. But I still count not define the issue.

[je-ik]

What you describe sounds like a deadlock. That was what I was trying to figure out, the metrics are generated in the checkpoint thread, but should be read in the thread that calls advance(). Therefore the backlog metrics should be accessible, the fact that they are not are quite suspicious. But other than that I currently don't have enough understanding of how the metrics are binded between beam metrics and flink metrics, very quickly looking into the code, I do not see how the backlog metrics could get out of the KafkaUnboundedReader, because I'm struggling to find how the metrics are "registered" in any container for export.

[github-actions]

This pull request has been marked as stale due to 60 days of inactivity. It will be closed in 1 week if no further activity occurs. If you think that’s incorrect or this pull request requires a review, please simply write any comment. If closed, you can revive the PR at any time and @mention a reviewer or discuss it on the dev@beam.apache.org list. Thank you for your contributions.

[github-actions]

This pull request has been closed due to lack of activity. If you think that is incorrect, or the pull request requires review, you can revive the PR at any time.