[
https://issues.apache.org/jira/browse/FLINK-8532?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16596197#comment-16596197
]
ASF GitHub Bot commented on FLINK-8532:
---------------------------------------
Guibo-Pan commented on issue #6544: [FLINK-8532] [Streaming] modify
RebalancePartitioner to use a random partition as its first partition
URL: https://github.com/apache/flink/pull/6544#issuecomment-416913773
@StephanEwen Thanks for your suggestion. I've made an update to it.
For further optimization, the one-channel-only selector which always returns
just `0`, can be used instead of using `ShufflePartitioner` or
`RebalancePartitioner` where outputs have only one channel.
It seems that we have a `ForwardPartitioner` already which always returns 0.
And it could be an optimization in the process during building the job graph,
maybe in another new issue.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> RebalancePartitioner should use Random value for its first partition
> --------------------------------------------------------------------
>
> Key: FLINK-8532
> URL: https://issues.apache.org/jira/browse/FLINK-8532
> Project: Flink
> Issue Type: Improvement
> Components: DataStream API
> Reporter: Yuta Morisawa
> Assignee: Guibo Pan
> Priority: Major
> Labels: pull-request-available
>
> In some conditions, RebalancePartitioner doesn't balance data correctly
> because it use the same value for selecting next operators.
> RebalancePartitioner initializes its partition id using the same value in
> every threads, so it indeed balances data, but at one moment the amount of
> data in each operator is skew.
> Particularly, when the data rate of former operators is equal , data skew
> becomes severe.
>
>
> Example:
> Consider a simple operator chain.
> -> map1 -> rebalance -> map2 ->
> Each map operator(map1, map2) contains three subtasks(subtask 1, 2, 3, 4, 5,
> 6).
> map1 map2
> st1 st4
> st2 st5
> st3 st6
>
> At the beginning, every subtasks in map1 sends data to st4 in map2 because
> they use the same initial parition id.
> Next time the map1 receive data st1,2,3 send data to st5 because they
> increment its partition id when they processed former data.
> In my environment, it takes twice the time to process data when I use
> RebalancePartitioner as long as I use other partitioners(rescale, keyby).
>
> To solve this problem, in my opinion, RebalancePartitioner should use its own
> operator id for the initial value.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
