[
https://issues.apache.org/jira/browse/YARN-1445?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13918984#comment-13918984
]
Zhijie Shen commented on YARN-1445:
-----------------------------------
I've thought about postponing the unregistration success notification until the
application is at FINISHED. It seems impossible, because it will result in a
dead lock bellow
1. AM container is waiting for finishing unregistration to move on and exit;
2. Unregistration is waiting for RM notifying success;
3. RM is waiting for RMApp moving from FINISHING to FINISHED to return success;
4. RMApp is waiting for RMAppAttempt moving from FINISHING to FINISHED;
5. RMAppAttempt is waiting for AM container being finished.
Then, if we return a prior state to the client given the internal FINISHING,
and still return unregistration success when RMApp reaches FINISHING, client
will see, for example, RUNNING, while the registration is already successful.
The inconsistency here may result in some race condition for the process
relying on checking the final state.
For example, MR client will direct user to AM if the application is said not to
be in a final state. Then, it is possible that AM is unregistered, and RM tells
the client that the application is still running. When the client moves on to
contact AM, AM has proceeded and exited before being able to respond the client
request.
It seems that we cannot avoid splitting the user-faced state, and FINISHING can
map to a period of an application's life cycle, which is from unregistration to
process exit. [~jlowe] and [~jianhe], how do you think about it?
> Separate FINISHING and FINISHED state in YarnApplicationState
> -------------------------------------------------------------
>
> Key: YARN-1445
> URL: https://issues.apache.org/jira/browse/YARN-1445
> Project: Hadoop YARN
> Issue Type: Bug
> Reporter: Xuan Gong
> Assignee: Xuan Gong
> Attachments: YARN-1445.1.patch, YARN-1445.2.patch, YARN-1445.3.patch,
> YARN-1445.4.patch, YARN-1445.5.patch, YARN-1445.5.patch, YARN-1445.6.patch
>
>
> Today, we will transmit both RMAppState.FINISHING and RMAppState.FINISHED to
> YarnApplicationState.FINISHED.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
