[ https://issues.apache.org/jira/browse/KAFKA-14366?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17630855#comment-17630855 ]
Chetan commented on KAFKA-14366: -------------------------------- Hi [~pnee] I will detail the steps to recreate the scenario in a day. I am currently collecting all details and will share the findings in detail. > Kafka consumer rebalance issue, offsets points back to very old committed > offset > -------------------------------------------------------------------------------- > > Key: KAFKA-14366 > URL: https://issues.apache.org/jira/browse/KAFKA-14366 > Project: Kafka > Issue Type: Bug > Components: consumer, offset manager > Affects Versions: 2.8.1 > Environment: Production > Reporter: Chetan > Priority: Major > Attachments: rebalance issue.docx > > > Hi All, > We are facing an issue while the client consumer restart (again not all > restarts are ending up with this issue) and during the re-balancing scenario, > sometimes one of the partition offsets goes back a long way from the > committed offset. > Scenario : > Assume we have 4 instances of consumer and restarts of consumer one after the > other. > # At the time of starting restarts assume the offset on partition 10 of a > topic being consumed is pointing to 50000. (last offset of the topic and 0 > lag) > # When restarts start (rebalancing) suddenly the offsets start pointing to > 20000. > # While all the restarts are going on the consumer who is attached starts > reading from 20000 and goes on. > # Once all rebalance is completed, and all messages from 20000 to 50000 > offset has been read (where it had stopped initially) > We end up having around 30K duplicates. > (The numbers here are just an example, in production, we are facing huge > duplicates and every two rebalance during restarts of consumer out of 10 > restart exercise activity ends up in such duplicates and not all partitions > and only one or two partitions behave this way and randomly) > This seems to be a bug. I am attaching all screenshots for reference as well. > Can someone kindly help out here? -- This message was sent by Atlassian Jira (v8.20.10#820010)