[
https://issues.apache.org/jira/browse/AURORA-582?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14065299#comment-14065299
]
Kevin Burg commented on AURORA-582:
-----------------------------------
I've worked out a fix for this at
https://github.com/foursquare/incubator-aurora/commit/50b1d197f4faaf0f770cee0552d88c2db0eea4a6.
This works in my own testing.
It seems that everywhere SchedulingFilter.filter is called, a resource object
is computed directly from an offer and passed into that function, however the
attribute object is not. The available attributes are later computed from a
slave name. My change pulls the attributes right out of the offer that the
resources are read from and doesn't go through AttributeStore. Is there a
reason why attributes were being retrieved in a much more difficult way than
resources?
I haven't been able to rewrite all the test cases, so those are failing, and I
would also like someone to check that I am doing to right thing in Preemptor.
> Slave attributes cached incorrectly by scheduler
> ------------------------------------------------
>
> Key: AURORA-582
> URL: https://issues.apache.org/jira/browse/AURORA-582
> Project: Aurora
> Issue Type: Bug
> Components: Scheduler
> Affects Versions: 1.0.0
> Reporter: Kevin Burg
> Assignee: Bill Farner
> Priority: Critical
>
> I would like to be able to update the attributes being offered by a mesos
> slave after the cluster has been setup. The cluster was initialized with two
> slaves reporting two attributes - 'host' and 'rack'. I was able to assign
> task with constraints involving these two attributes and everything works.
> Any third attribute I add shows up in the '/offers' endpoint of the aurora
> process, but not in the '/slaves' endpoint. Modifications to existing
> attributes are similarly not reflected. I am not able to schedule tasks
> involving a constraint on the third attribute or on any modified attributes.
> The schedule requests just end in vetoes from both slaves unable to resolve
> the constraints.
> It's been suggested that I delete the slave's 'meta' directory before
> bouncing the slave, but this has not helped to solve the issue. The only way
> I have been able to modify the attributes so that aurora can see them is by
> deleting the entire AURORA_ROOT and rebuilding the cluster.
--
This message was sent by Atlassian JIRA
(v6.2#6252)