[
https://issues.apache.org/jira/browse/AURORA-413?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13995853#comment-13995853
]
Bill Farner commented on AURORA-413:
------------------------------------
To give another angle on Maxim's reply, the default behavior of the updater is
all-or-none. If the desired end state cannot be fully achieved, the update is
considered a failure.
> aurora update fails if update results in a pending job
> ------------------------------------------------------
>
> Key: AURORA-413
> URL: https://issues.apache.org/jira/browse/AURORA-413
> Project: Aurora
> Issue Type: Bug
> Components: Client
> Affects Versions: 0.5.0
> Reporter: Anindya Sinha
>
> Assume I have a job running on the cluster with 2 instances. Let us say both
> tasks are in RUNNING state.
> At this point, I update my job to bump up the # of instances from 2 to 4, and
> do an aurora update. As expected, it leaves the 2 running instances intact,
> and attempts to start instances 3 and 4. Assume instance 3 starts up fine and
> is in RUNNING state. If for some reason, the cluster is in such a state that
> instance 4 cannot be scheduled immediately (ie. it is in PENDING state). In
> that case, the update fails if the task does not move to RUNNING state in
> 30-45 seconds after attempting to launch. Since update fails, it kills
> instance 3 as well.
> aurora update should not fail if a task is in PENDING state since that task
> may get scheduled when some other job finishes or fails. Also, if I were to
> do a aurora create instead with # of instances = 4, it keeps 3 of them in
> RUNNING while the 4th instance is in PENDING state. So, the behavior is
> different depending on whether we do a "aurora create" v "aurora update"
> which ideally should not be the case.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
