Re: Telling if a job has caught up with Kafka
Posted by
elmosca on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Telling-if-a-job-has-caught-up-with-Kafka-tp12261p12263.html
Hi,
We are interested on this too. So far we flag the records with timestamps in different points of the pipeline and use metrics gauges to measure latency between the different components, but would be good to know if there is something more specific to Kafka that we can do out of the box in Flink.
Cheers,
Bruno
Hi,
thank you Gyula for posting that question. I’d also be interested in how this could be done.
You mentioned the dependency on the commit frequency. I’m using https://github.com/quantifind/KafkaOffsetMonitor. With the 08 Kafka consumer a job's offsets as shown in the diagrams updated a lot more regularly than the checkpointing interval. With the 10 consumer a commit is only made after a successful checkpoint (or so it seems).
Why is that so? The checkpoint contains the Kafka offset and would be able to start reading wherever it left off, regardless of any offset stored in Kafka or Zookeeper. Why is the offset not committed regularly, independently from the checkpointing? Or did I misconfigure anything?
Thanks
Florian
> Am 17.03.2017 um 10:26 schrieb Gyula Fóra <[hidden email]>:
>
> Hi All,
>
> I am wondering if anyone has some nice suggestions on what would be the simplest/best way of telling if a job is caught up with the Kafka input.
> An alternative question would be how to tell if a job is caught up to another job reading from the same topic.
>
> The first thing that comes to my mind is looking at the offsets Flink commits to Kafka. However this will only work if every job uses a different group id and even then it is not very reliable depending on the commit frequency.
>
> The use case I am trying to solve is fault tolerant update of a job, by taking a savepoint for job1 starting job2 from the savepoint, waiting until it catches up and then killing job1.
>
> Thanks for your input!
> Gyula