[
https://issues.apache.org/jira/browse/FLINK-13165?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16881278#comment-16881278
]
Zhu Zhu edited comment on FLINK-13165 at 7/9/19 2:58 PM:
---------------------------------------------------------
Hi Till, just agree this mechanism could reduce the chance that "consumer task
gets deployed before a producer causing a resource deadlock situation".
But seems this may not guarantee producer scheduled before consumer if
InputDependencyConstraint=ANY. e.g. topology \{A->D, B->C->D}, all blocking
edges, A,B scheduled first, then A finishes and start scheduling D, this
happens before C get scheduled.
But if we set InputDependencyConstraint=ALL, there is no chance that the
consumer get scheduled before its producers. Thus the resource deadlock can be
avoided even without the mechanism introduced in this JIRA.
was (Author: zhuzh):
Agree this mechanism could reduce the chance that "consumer task gets deployed
before a producer causing a resource deadlock situation".
But seems to me this may not guarantee producer scheduled before consumer if
InputDependencyConstraint=ANY. e.g. topology \{A->D, B->C->D}, all blocking
edges, A,B scheduled first, then A finishes and start scheduling D, this
happens before C get scheduled.
But if we set InputDependencyConstraint=ALL, there is no chance that the
consumer get scheduled before its producers. Thus the resource deadlock can be
avoided even without the mechanism introduced in this JIRA.
> Complete requested slots in request order
> -----------------------------------------
>
> Key: FLINK-13165
> URL: https://issues.apache.org/jira/browse/FLINK-13165
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Coordination
> Affects Versions: 1.9.0
> Reporter: Till Rohrmann
> Assignee: Till Rohrmann
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.9.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> When executing batch jobs with fewer slots than requested we should make sure
> that the slot requests are being completed in the order in which they were
> enqueued into the {{SlotPool}}. Otherwise we might risk that a consumer task
> gets deployed before a producer causing a resource deadlock situation.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
