Closed gfalcone closed 4 months ago
Also, it might be good to add the --attached
option to this page : https://beam.apache.org/documentation/runners/flink/
there wasn't much changes for flink runner between Beam 2.51.0 and 2.52.0 except for #28614, which known to have caused a couple of flaky tests and then fixed later.
Is attached mode works fine?
Impossible to test on my end with the setup I have. When running through FlinkDeployment
Kubernetes object, it is submitted in detached mode...
Similar thing happening with AWS SqsIO after update to 2.52.
It seems that #28614 replaced UnboundedSourceWrapper
with FlinkUnboundedSource
and WatermarkStrategy.noWatermarks()
, effectively blocking watermark emission.
CC: @jto who has been quite responsive. Would appreciate if you can share some ideas about fixing https://github.com/apache/beam/issues/29902#issuecomment-1883597256
Move to P1 as this essentially breaking streaming FlinkRunner. If it's not fixed in the next release I would suggest revert #28614 and related changes.
@noster-dev @gfalcone are you able to test that reverting that change would resolve the issue?
I can confirm that reverting to 2.51 fixes the problem. Tomorrow I will try to test it on 2.52 with #28614 reverted
Hi, are there any updates to this issue? I'm facing a potentially similar issue with Kafka source not committing offsets back even though commitsOffsetsInFinalize is true and checkpointing is enabled. Beam v2.52.0. Flink v1.12.4.
Committing offsets works in Beam v2.51.0.
Hey there! Apologies for not looking into this earlier. I was on parental leave and had a lot to catch up when I got back. I think this issue and https://github.com/apache/beam/issues/30539 are duplicates. Taking a look now.
This is very much likely related to #25525 as after the change to FLIP-27 sources call to finalizeCheckpoint
seem to disappear. The source refactor likely caused #29816 as well, though the specific mechanics are still unknown.
Hey there! Apologies for not looking into this earlier. I was on parental leave and had a lot to catch up when I got back. I think this issue and #30539 are duplicates. Taking a look now.
Yes, also related to missing finalizeCheckpoint
call.
What happened?
Hello !
I have a streaming job processing messages from Pub/Sub that does not work anymore using Beam 2.52.0 with Flink Runner (in detached mode)
The pipeline works fine in Beam 2.51.0
Here is the code of the pipeline :
With the pom.xml :
And the associated configuration :
Here is the screenshot from Google Cloud Console showing that messages are never acked :
Thank you for your help :)
Issue Priority
Priority: 2 (default / most bugs should be filed as P2)
Issue Components