Duplicated messages when kafka-monitor is restarted about kafka-monitor

[jlisam]

An update on the above, I restarted this again but this time the indices make sense. The consumer picks up the latest produced index.

[2018-08-03 00:38:32,485] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=4 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,486] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=7 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,486] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=0 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,487] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=6 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,487] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=12 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,488] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=8 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,488] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=15 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,488] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=14 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,490] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=3 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,491] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=11 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,493] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=5 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,493] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=13 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,500] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=1 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,500] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=9 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,513] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=2 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,514] INFO single-cluster-monitor/Current index=94, nextIndex=-1 for partition=10 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,586] INFO single-cluster-monitor/Current index=95, nextIndex=95 for partition=7 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,587] INFO single-cluster-monitor/Current index=95, nextIndex=95 for partition=4 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,587] INFO single-cluster-monitor/Current index=95, nextIndex=95 for partition=0 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,588] INFO single-cluster-monitor/Current index=95, nextIndex=95 for partition=6 (com.linkedin.kmf.services.ConsumeService)
[2018-08-03 00:38:32,588] INFO single-cluster-monitor/Current index=95, nextIndex=95 for partition=15 (com.linkedin.kmf.services.ConsumeService)

[lindong28]

Hey @jlisam,

Thanks for raising the issue. Here is the expected behavior of after kafka-monitor is restarted:

• For any given partition, ProduceService will send a message with index = 0. Say the offset of this produced message is offset1. Then the producer will send messages with index sequence (offset1, offset1+1, offset1+2,...). So the sequence of indexes of messages from ProduceService is (0, offset1, offset1 + 1, ...)

• For any given partition, ConsumeService starts consuming from the log end offset, which means that ConsumeService will not consume any message that was produced before it is started. Say the sequence of indexes of the messages it has consumed is (m1, m2, m3, ...). ConsumeService will discard the first offset m2 and start from the second offset. It expects offsets starting from the second offset to be consecutive. In other words, it expect sequence of indexes to be (m1, m2, m2 + 1, m2 + 2, ..). If there is a message whose index is smaller than the previous index (not including the first message), then this is considered a message duplication.

So it looks like the logic is correct for kafka-monitor restart. I think we can prove that, if there is no message loss and message duplication in the Kafka side, ConsumeService should always see consecutively incrementing indexes, right?

Given that you actually observed something weird, there may be problem somewhere. I currently don't have an explanation for that.

[jlisam]

thanks @lindong28 for the prompt response! I will try to track this. So far I can't reproduce this locally but when my instance is running in kubernetes, I almost always hit that case.

[lindong28]

@jlisam Thanks for reporting this issue! Since LinkedIn Kafka team does not use Kubernetes, it is hard for us to investigate this issue if this issue only happens when Kafka Monitor is used with Kubernetes. We will try to fix the issue if you or other use can provide more specific information on what is going wrong with Kafka Monitor to help us debug it.

[lindong28]

I will close this issue. Please feel free to re-open if there is more actionable information for us.