[jira] [Commented] (FLINK-8532) RebalancePartitioner should use Random value for its first partition

["Guibo Pan (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-8532?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16577437#comment-16577437 ] 

Guibo Pan commented on FLINK-8532:
----------------------------------

[~Morisawa] In my case, it becomes a big problem when the parallelism of the next operator grows.

For example, in my job, the parallelism of the next operator is more than 10k, it costs several minutes to travel through all the outputChannels. 

Messages are send to the 0th outputChannel at one moment so that causes lag of the 0th subtask,  and then the 1st outputChannel next moment, and so on. At  every moment, some of the subtask are busy processing messages with "big" lag, while the others are idle with no message arrived.

And this may make the delay becomes twice or more for one message.

 

I would like to work on this issue, however RebalancePartitioner doesn't have some id like "operator id", I think it can start with a random partition simply.

 

> 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
>            Priority: Minor
>
> 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)