[
https://issues.apache.org/jira/browse/BEAM-4803?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16576566#comment-16576566
]
Raghu Angadi commented on BEAM-4803:
------------------------------------
It is not configurable. It was chosen mainly as a 'reasonable value' :). In
most of pipelines, Dataflow bundles are small enough that split gets scheduled
very often. Spurious closes due to 1 minute timeout is not very common, it can
happen if there is too much data buffered downstream. There was one issue
reported on Stackoverflow where user had set up such that Kafka metrics were
reported with a unique id for each instance of KafkaClient :
[https://stackoverflow.com/questions/48765660/google-dataflow-idle-kafka-readers]
.
> Beam spark runner not working properly with kafka
> -------------------------------------------------
>
> Key: BEAM-4803
> URL: https://issues.apache.org/jira/browse/BEAM-4803
> Project: Beam
> Issue Type: Bug
> Components: io-java-kafka, runner-spark
> Affects Versions: 2.4.0, 2.5.0
> Reporter: Vivek Agarwal
> Assignee: Jean-Baptiste Onofré
> Priority: Major
>
> We are running a beam stream processing job on a spark runner, which reads
> from a kafka topic using kerberos authentication. We are using java-io-kafka
> v2.4.0 to read from kafka topic in the pipeline. The issue is that the
> kafkaIO client is continuously creating a new kafka consumer with specified
> config, doing kerberos login every time. Also, there are spark streaming jobs
> which get spawned for the unbounded source, every second or so even when
> there is no data in the kafka topic. Log has these jobs-
> INFO SparkContext: Starting job: dstr...@sparkunboumdedsource.java:172
> We can see in the logs
> INFO MicrobatchSource: No cached reader found for split:
> [org.apache.beam.sdk.io.kafka.KafkaUnboundedSource@2919a728]. Creating new
> reader at checkpoint mark...
> And then it creates new consumer doing fresh kerberos login, which is
> creating issues.
> We are unsure of what should be correct behavior here and why so many spark
> streaming jobs are getting created. We tried the beam code with flink runner
> and did not find this issue there. Can someone point to the correct settings
> for using unbounded kafka source with spark runner using beam?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
