[
https://issues.apache.org/jira/browse/TEZ-900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14304269#comment-14304269
]
Chang Li commented on TEZ-900:
------------------------------
[~hitesh] YarnClientImpl's check isn't complete. The "failure of submit to the
wrong queue" doesn't really fail to be submitted, instead it succeed in being
submitted, however failed immediately after. The check in YarnClientImpl is
simply
"if (!state.equals(YarnApplicationState.NEW) &&
!state.equals(YarnApplicationState.NEW_SAVING)) {
LOG.info("Submitted application " + applicationId);
break;
}"
because the state at that time is now FAILED, it will break out the loop and
not be captured within YarnClient. But YarnRunner did a more thorough check, it
did check FAILED and KILLED state immediate after launch,
"if (appMaster == null
|| appMaster.getYarnApplicationState() == YarnApplicationState.FAILED
|| appMaster.getYarnApplicationState() ==
YarnApplicationState.KILLED) {
throw new IOException("Failed to run job : " +
diagnostics);
}"
I can open a jira on yarn for this issue, and I think the solution is simply
duplicating the check logic from YarnRunner to YarnClient
> Confusing message for incorrect queue for some tez examples
> -----------------------------------------------------------
>
> Key: TEZ-900
> URL: https://issues.apache.org/jira/browse/TEZ-900
> Project: Apache Tez
> Issue Type: Bug
> Affects Versions: 0.3.0
> Reporter: Jonathan Eagles
> Assignee: Chang Li
> Attachments: tez900.patch
>
>
> The error message of at least orderedwordcount with non-existent queue is
> very difficult as compared to mrrsleep which is very intuitive.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
