[
https://issues.apache.org/jira/browse/FLINK-14510?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17135674#comment-17135674
]
Till Rohrmann commented on FLINK-14510:
---------------------------------------
This sounds good. Just as a side comment, we might need to make the
{{ExecutionGraph}} construction lazy again in the future. This might be
necessary in order to support the reactive mode where the set of resources
might change over time. Consequently, one might be able to run a stage with a
larger/smaller parallelism than one initially thought.
> Remove the lazy vertex attaching mechanism from ExecutionGraph
> --------------------------------------------------------------
>
> Key: FLINK-14510
> URL: https://issues.apache.org/jira/browse/FLINK-14510
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Coordination
> Affects Versions: 1.10.0
> Reporter: Zhu Zhu
> Assignee: Zhu Zhu
> Priority: Major
> Fix For: 1.12.0
>
>
> Currently in production, the vertex attaching is only invoked right after the
> ExecutionGraph is created in ExecutionGraphBuilder. That means lazy attaching
> is not necessary at the moment. It however adds extra complexity to
> ExecutionGraph, since we need to assume that the vertices may be not
> initialized or even get changed.
> Moreover, attaching vertices after a job starts scheduling is an undefined
> behavior which would not work properly.
> I'd propose to remove the lazy attaching mechanism, and do vertices building
> and related components initialization in ExecutionGraph constructor.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
