[
https://issues.apache.org/jira/browse/AURORA-1075?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14303775#comment-14303775
]
Bill Farner commented on AURORA-1075:
-------------------------------------
> If instances scheduled in this manner are required to run on every slave
> (such as a logger, or some other utility service) then they are effectively
> immune from preemption.
Or perhaps we need to reintroduce a priority spectrum (we had this at one
time), with a range that only administrators can use.
On the flip side, though, we're currently not set up to preempt tasks to make
space for this type of job on a host. Definitely doable, but requires some
work on the preemptor.
> Could you elaborate on some of the pitfalls such a feature might introduce in
> a multi-framework cluster?
Preemption is the biggest inhibitor. Until Aurora has visibility into
resources in use by other frameworks, and the means to terminate them, our
hands would be tied to make the kind of guarantee needed by this type of
feature. Of course, you are fine today if you configure mesos to statically
partition the cluster. I believe MESOS-1607 is the most relevant ticket on the
mesos side to support this.
> schedule instance on every host matching constraints
> ----------------------------------------------------
>
> Key: AURORA-1075
> URL: https://issues.apache.org/jira/browse/AURORA-1075
> Project: Aurora
> Issue Type: Story
> Components: Scheduler
> Reporter: Jay Buffington
>
> I'd like aurora to schedule an instance of a given job on every host it has
> seen an offer for. So if I were to create this job:
> {noformat}
> import os
> jobs = [Job(
> task=SimpleTask(name="hello_world", command="echo hello world"),
> role=os.getenv('USER'),
> cluster="cluster1",
> instances="*"
> )]
> {noformat}
> It would result in hello world running on every host in the cluster where
> there is room.
> If I were to set constraints (e.g. dedicated role), only hosts matching
> constraints would have the instances started on it.
> If the Job is a production job, instances of other jobs would be preempted if
> necessary to make room for this instance on every host aurora knows about.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
