[
https://issues.apache.org/jira/browse/MAPREDUCE-5576?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13792744#comment-13792744
]
Zhijie Shen commented on MAPREDUCE-5576:
----------------------------------------
bq. IIUC, failing the unregistration fails the application.
Yes, the current design is to fail the application when unregistration is
failed. What I mean is that error of getting the history url should not fail
unregistration, because ApplicationMasterProtocol#finishApplicationMaster
doesn't require to url to finis an application on RM side. The error of getting
the history url may be caused by unavailability of JHS, but also be caused by
mis configuration (actually we ran into this issue). JHS HA will reduce the
chance of error, but it will still happen.
bq. If we decide to still go through with this change, we should probably fail
the application early - before running the job and not after.
Agree. In fact, when the url is not available, RM already has the logic to log
something. Maybe we want to enhance the log.
bq. A better approach might to be explicitly log or show that the JHS is
down/inaccessible/mis-configured and setting the tracking URL has failed.
I don't want to fail the application. As is mentioned above, I don't think it
make sense that error of getting the history url should fail unregistration,
and ultimately fail the application.
> MR AM unregistration should be failed due to UnknownHostException on getting
> history url
> ----------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-5576
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-5576
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Zhijie Shen
> Assignee: Zhijie Shen
>
> Before RMCommunicator sends the request to RM to finish the application, it
> will try to get the JHS url, which may throw UnknownHostException. The
> current code path will skip sending the request to RM when the exception is
> raised, which sounds not a reasonable behavior, because RM's unregistering an
> AM will not affected by the tracking URL. The URL can be empty or null.
> AFAIK, the impact of null URL will be that the URL to redirect users from RM
> web page to JHS will be unavailable, and the job report will not show the URL
> as well. However, is it much much better than failing an application because
> of UnknownHostException here? Anyway, users can go to JHS directly to find
> the application history info.
> Therefore, the reasonable code path here should be catching
> UnknownHostException and set historyUrl = null
--
This message was sent by Atlassian JIRA
(v6.1#6144)
