[
https://issues.apache.org/jira/browse/OOZIE-3646?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17458267#comment-17458267
]
Dénes Bodó commented on OOZIE-3646:
-----------------------------------
[~zuston] Thanks for catching this. I didn't take a deeper look but could you
please also provide unit tests for this change. A test which can result in
deadlock with the old code but passes with the new code would be great.
Thanks
> Possible dead-lock in SignalXCommand
> ------------------------------------
>
> Key: OOZIE-3646
> URL: https://issues.apache.org/jira/browse/OOZIE-3646
> Project: Oozie
> Issue Type: Bug
> Reporter: Junfan Zhang
> Priority: Major
>
> The limited thread execution mechanism aims to solve the dead-lock when all
> active threads are executing the SignalXCommand's invokeAll method.
> h2. Dead-lock when to happen
> Assuming that Oozie CallableQueue thread pool size is 120, when all threads
> are executing the {{SignalXCommand.startForkedActions}} method, a deadlock
> occurs.
> Because in {{SignalXCommand.startForkedActions}}, the code of
> {{List<Future<ActionExecutorContext>> futures =
> Services.get().get(CallableQueueService.class)
> .invokeAll(tasks);}} will be sync executed, however now
> all callableQueue threads are busy.
> h2. Solution
> 1. Need to limit directly invokeAll call when the num of rest threads is less
> than the tasks
> 2. To obtain correct active threads number in callableQueue, the
> SignalXCommand.class lock is needed.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
