[
https://issues.apache.org/jira/browse/FLINK-14051?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16931275#comment-16931275
]
TisonKun commented on FLINK-14051:
----------------------------------
Thanks for your explanation [~till.rohrmann].
I think the problem is the approach we compile Flink job. In per-job(detached)
mode we use {{OptimizerPlanEnvironment}} which abort execution and set the
FlinkPlan to be executed. Thus when the job consists of multiple parts it
always surprises users that the following parts are never executed. We deploy a
session cluster in attached mode to workaround this a bit.
Maybe standard what per-job stands for could help. In my opinion per-job is
like Driver mode in Spark scope that we just submit jars represent user job to
the resource management cluster(YARN, Mesos, k8s) and a ClusterEntrypoint is
launched and then compile to get the JobGraph and start to execute it. With
this definition we survive from the implementation details for compiling
JobGraph in client side.
> Deploy job cluster in attached mode
> -----------------------------------
>
> Key: FLINK-14051
> URL: https://issues.apache.org/jira/browse/FLINK-14051
> Project: Flink
> Issue Type: Sub-task
> Components: Client / Job Submission, Command Line Client
> Affects Versions: 1.10.0
> Reporter: TisonKun
> Priority: Major
> Fix For: 1.10.0
>
>
> While working on FLINK-14048 I revisit the problem we handle deploy logic in
> a complicated if-else branches in {{CliFrontend#runProgram}}. Previously we
> said even in per-job mode and attached we deploy a session cluster for
> historical reasons.
> However, I notice that {{#deployJobCluster}} has a parameter {{boolean
> detached}}. Also it is used in sql-client package. So it looks like we can
> deploy job cluster in attached mode as we do in sql-client package.
> However, as [~xccui] answered on mailing list
> [here|https://lists.apache.org/x/thread.html/5464459db08f2a756af0c61eb02d34a26f04c27c62140886cad52731@%3Cuser.flink.apache.org%3E],
> we support only standalone session cluster for sql-client. So maybe it is
> not our case. Anyway, if we cannot deploy job cluster in attached mode, I'd
> like to know the concrete reason.
> CC [~till.rohrmann] [~twalthr]
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
