[
https://issues.apache.org/jira/browse/OOZIE-2345?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14725688#comment-14725688
]
Purshotam Shah commented on OOZIE-2345:
---------------------------------------
Approach is to have a new command ForkedActionStartXCommand, which will acquire
lock on wf action. It will just submit job and won't do any wf update. All Wf
update will be done by signalXcommand.
SignalXcommand will submit all ForkedActionStartXCommand and wait for them to
complete.
If any ForkedActionStartXCommand has failed, SignalXcommand will fail the
workflow.
In case of transient or user retry. ActionStartXcommand will be queued, which
need to acquire lock of Wf.
This feature can be enable/disable using
{{oozie.workflow.parallel.fork.action.start}}
> Concurrent job submission for forked action
> -------------------------------------------
>
> Key: OOZIE-2345
> URL: https://issues.apache.org/jira/browse/OOZIE-2345
> Project: Oozie
> Issue Type: Bug
> Reporter: Purshotam Shah
> Assignee: Purshotam Shah
> Attachments: OOZIE-2345-V3.patch
>
>
> We have few customers whose SLA is 8 min. They have around 30 actions. There
> are 25 actions in fork.
> Though forked action jobs runs concurrently. forked action job submission is
> sequential.
> Whenever NN is slow, job submission takes more time. Even if job submission
> is delay for 30 sec. Total WF delay will be ~12 min.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
