[
https://issues.apache.org/jira/browse/FLINK-23402?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17383090#comment-17383090
]
Till Rohrmann commented on FLINK-23402:
---------------------------------------
+1 for unifying the way the exchange mode is configured. I also really like
Stephan's proposal for the 3 different batch modes. I do agree that they make a
lot more sense than what we currently have and will make it easier for our
users to use and understand Flink.
Scoping wise implementing first
{{RuntimeMode.BATCH.withAllExchangesBlocking()}} and
{{RuntimeMode.BATCH.willAllExchangesPipelined()}} could be a start. However we
should consider that changing the default from {{withAllExchangesBlocking()}}
to pipelined within slot sharing groups if it's a {{FORWARD}} edge, can be a
behavioural change in terms of resource consumption. Consequently, this might
break some jobs that all of a sudden might need more resources for a slot.
> 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
> Priority: Major
>
> 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.
> It would include the following steps:
> - Move {{GlobalDataExchangeMode}} from {{o.a.f.streaming.api.graph}} to
> {{o.a.f.api.common}} (with reworked JavaDocs) as {{ExchangeMode}} (to have it
> shorter) next to {{RuntimeMode}}
> - Add {{StreamExecutionEnvironment.setExchangeMode()}} next to
> {{setRuntimeMode}}
> - Add option {{execution.exchange-mode}}
> - Add checks for invalid combinations to StreamGraphGenerator
> - Deprecate ExecutionMode ([avoid
> confusion|https://stackoverflow.com/questions/68335472/what-is-difference-in-runtimeexecutionmode-and-executionmode])
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
