[
https://issues.apache.org/jira/browse/FLINK-31065?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17689066#comment-17689066
]
luoyuxia commented on FLINK-31065:
----------------------------------
[~yunfanfight...@foxmail.com] Thanks for raising it up. I also noticed it
before and then tried a static split assign in our internal TPC-DS benmark.
Unfortunately, I haven't found any performance improvement, so I give up. But I
do think it's a good point considering failover.
Apart from introducing static split assign strategy, will you plan also to make
some sources like Hive use it?
> Support more split assigner strategy for batch job
> --------------------------------------------------
>
> Key: FLINK-31065
> URL: https://issues.apache.org/jira/browse/FLINK-31065
> Project: Flink
> Issue Type: Improvement
> Reporter: yunfan
> Priority: Major
>
> Currently flink use LocatableInputSplitAssigner as the default split
> assigner.
> Which splits the task will consume are dynamic in the runtime.
> It is not a good strategy in the batch mode.
> For example, we have 100 splits and the job has 100 tasks.
> When the job start, we don't have enough resource to start the 100 tasks,
> only 10 tasks started in the first time.( it is a common case in batch mode)
> These 10 tasks will consume all splits, and other 90 tasks will reach finish
> state.
> This is obviously not a good idea in batch mode.
> In extreme cases, 99 tasks finished, only one task running, and if this
> running task failed,
> it will take much time to rerun this task( Because it need to consume it's 10
> split again).
> Spark will bind splits to it's task, I think it is a better way in the batch
> mode.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
