[
https://issues.apache.org/jira/browse/AURORA-1111?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14310372#comment-14310372
]
Jay Buffington commented on AURORA-1111:
----------------------------------------
While this is independent of AURORA-1075, it will likely be used in tandem with
it. With AURORA-1075, you'll be able to run an instance on every slave in the
cluster. If you do that with production=True you'll need your quota to scale
linearly with the number of slaves in the cluster. Plus, if a slave is full of
production jobs, we need to be able to preempt one of those to make room for a
cluster service.
> support a cluster service job priority
> ---------------------------------------
>
> Key: AURORA-1111
> URL: https://issues.apache.org/jira/browse/AURORA-1111
> Project: Aurora
> Issue Type: Story
> Components: Scheduler
> Reporter: Jay Buffington
>
> I propose we support a "cluster service" (I'm open to alternative names)
> priority level which is higher than production. Jobs in this priority would
> be able to preempt production jobs and would not be subject to quota. Only a
> very small set of cluster administrators would be able to start jobs at this
> priority level.
> Thermos observer in it's current incarnation is the best example of something
> that would run at this priority. We currently accomplish guaranteeing that
> the observer is running by starting it out of band of mesos. With this
> change we would be able to start an observer on a slave through aurora, even
> if that particular slave is "full" of production tasks.
> There are other examples of cluster services, such as a splunk forwarder, or
> the mesos-slave process (eek! turtles!)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
