[
https://issues.apache.org/jira/browse/FLINK-26548?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17507075#comment-17507075
]
Zhu Zhu edited comment on FLINK-26548 at 3/15/22, 5:19 PM:
-----------------------------------------------------------
As discussed in
[#19040|https://github.com/apache/flink/pull/19040#issuecomment-1067723796], a
proper fix of this problem would be to rework
`StreamExecutionEnvironment#createFileInput(...)` to use new sources(FLIP-27).
We would mark this problem as a known issue of AdaptiveBatchScheduler and
postpone the fix to 1.16 (possibility will back port it to 1.15.1).
was (Author: zhuzh):
As discussed in #19040, a proper fix of this problem would be to rework
`StreamExecutionEnvironment#createFileInput(...)` to use new sources(FLIP-27).
We would mark this problem as a known issue of AdaptiveBatchScheduler and
postpone the fix to 1.16 (possibility will back port it to 1.15.1).
> the source parallelism is not set correctly with AdaptiveBatchScheduler
> -----------------------------------------------------------------------
>
> Key: FLINK-26548
> URL: https://issues.apache.org/jira/browse/FLINK-26548
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Task
> Affects Versions: 1.15.0
> Reporter: zl
> Assignee: zl
> Priority: Critical
> Labels: pull-request-available
> Fix For: 1.16.0
>
> Attachments: image-2022-03-09-19-00-18-396.png
>
>
> When running *_org.apache.flink.table.tpcds.TpcdsTestProgram_* with
> {_}*AdaptiveBatchScheduler*{_}, I ran into a problem：the num of records sent
> by the source operator is always 1, and the parallelism of source operator is
> also 1 even I set
> *_jobmanager.adaptive-batch-scheduler.default-source-parallelism_* to 8.
> !image-2022-03-09-19-00-18-396.png!
> After some research, I found that the operator A is not the actual file
> reader, it just splits files and assigns splits to downstream tasks for
> further processing, and the operator B is the actual file reader task. Here,
> the parallelism of operator B is 64, and the records sent by operator A is 1,
> this means, operator A assigned all splits to a task of operator B, {*}_the
> other 63 tasks of operator B is idle_{*}, it is unreasonable.
> In this case, the parallelism of operator B should be
> *_jobmanager.adaptive-batch-scheduler.default-source-parallelism_* and the
> num of records sent by operator A also should be
> {*}_jobmanager.adaptive-batch-scheduler.default-source-parallelism_{*}.
>
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
