[ 
https://issues.apache.org/jira/browse/STORM-1838?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15289405#comment-15289405
 ] 

ASF GitHub Bot commented on STORM-1838:
---------------------------------------

Github user hmcl commented on the pull request:

    https://github.com/apache/storm/pull/1419#issuecomment-220102668
  
    @flisky can you please provide specific examples or a code snippet that I 
can use to test what you are saying. Can you also post the logs (perhaps on 
JIRA) that show the cases that are failing? What do you mean by `offset out of 
range`?
    
    As for your first comment, please keep in mind that the 
`FirstPollOffsetStrategy`, as the name indicates, it's only for the first poll, 
i.e., the first time a `KafkaSpout` polls from Kafka. It does not apply to 
consumer rebalance. Quoting the [javadoc](https://goo.gl/PyR7Cm) of "The offset 
used by the Kafka spout in the first poll to Kafka broker." Therefore, the 
`LATEST/EARLIEST` strategies only apply in the first poll, not on every 
consumer rebalance. Furthermore, if a consumer rebalance occurs, but the same 
partition is assigned to this spout, and some offsets have already been emitted 
and acked, we wan to keep those, and only process the ones that haven't yet 
been acked because they have have failed and waiting to be retried (according 
to exponential backoff) or may be on the "wire". 
    
    What do you mean by -2 , -1? Can you please give specific examples. Thanks.


> [storm-kafka-client] the resumed OffsetEntry goes different
> -----------------------------------------------------------
>
>                 Key: STORM-1838
>                 URL: https://issues.apache.org/jira/browse/STORM-1838
>             Project: Apache Storm
>          Issue Type: Bug
>            Reporter: Jifeng Yin
>             Fix For: 1.0.1
>
>
> There are no more new messages after consumer rejoins the group. And it turns 
> out the old OffsetEntry instance kept commitOffsetsForAckedTuples failing.
> the comment "leave the acked offsets as they were to resume where it left 
> off" doesn't work for me, but the ackedMsgs goes different.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to