[
https://issues.apache.org/jira/browse/FLINK-23402?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17391383#comment-17391383
]
Timo Walther commented on FLINK-23402:
--------------------------------------
Thanks for the feedback everyone. I will reopen this issue and implement the
mentioned changes:
- {{ShuffleMode}} -> {{BatchShuffleMode}}
- Drop {{AUTOMATIC}} and make {{ALL_EXCHANGES_BLOCKING}} the default value
- Rename option key {{execution.shuffle-mode}} to
{{execution.batch-shuffle-mode}}
> Expose a consistent GlobalDataExchangeMode
> ------------------------------------------
>
> Key: FLINK-23402
> URL: https://issues.apache.org/jira/browse/FLINK-23402
> Project: Flink
> Issue Type: Sub-task
> Components: API / DataStream
> Reporter: Timo Walther
> Assignee: Timo Walther
> Priority: Major
> Labels: pull-request-available
>
> The Table API makes the {{GlobalDataExchangeMode}} configurable via
> {{table.exec.shuffle-mode}}.
> In Table API batch mode the StreamGraph is configured with
> {{ALL_EDGES_BLOCKING}} and in DataStream API batch mode
> {{FORWARD_EDGES_PIPELINED}}.
> I would vote for unifying the exchange mode of both APIs so that complex SQL
> pipelines behave identical in {{StreamTableEnvironment}} and
> {{TableEnvironment}}. Also the feedback a got so far would make
> {{ALL_EDGES_BLOCKING}} a safer option to run pipelines successfully with
> limited resources.
> [~lzljs3620320]
> {quote}
> The previous history was like this:
> - The default value is pipeline, and we find that many times due to
> insufficient resources, the deployment will hang. And the typical use of
> batch jobs is small resources running large parallelisms, because in batch
> jobs, the granularity of failover is related to the amount of data processed
> by a single task. The smaller the amount of data, the faster the fault
> tolerance. So most of the scenarios are run with small resources and large
> parallelisms, little by little slowly running.
> - Later, we switched the default value to blocking. We found that the better
> blocking shuffle implementation would not slow down the running speed much.
> We tested tpc-ds and it took almost the same time.
> {quote}
> [~dwysakowicz]
> {quote}
> I don't see a problem with changing the default value for DataStream batch
> mode if you think ALL_EDGES_BLOCKING is the better default option.
> {quote}
> In any case, we should make this configurable for DataStream API users and
> make the specific Table API option obsolete.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
