[
https://issues.apache.org/jira/browse/MESOS-3912?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15006607#comment-15006607
]
Alexander Rukletsov commented on MESOS-3912:
--------------------------------------------
We want to *always* rescind after the capacity heuristic (sanity check). The
reason for this is the race between the allocator and the master: it can happen
that there are not enough free resources at allocator's disposal when it is
notified about the quota, but at this point it's too late to rescind. Once
offer management (including rescinding) is done in the allocator, this race is
gone and we should not rescind in the master.
The reason for the additional check is because we would like to make it
possible for each framework in the role for which quota is set to get an offer.
Though this is not strictly necessary, we think this will increase the
debugability and will improve user experience.
> Rescind offers in order to satisfy quota
> ----------------------------------------
>
> Key: MESOS-3912
> URL: https://issues.apache.org/jira/browse/MESOS-3912
> Project: Mesos
> Issue Type: Task
> Components: master
> Reporter: Alexander Rukletsov
> Assignee: Alexander Rukletsov
> Labels: mesosphere
>
> When a quota request comes in, we may need to rescind a certain amount of
> outstanding offers in order to satisfy it. Because resources are allocated in
> the allocator, there can be a race between rescinding and allocating. This
> race makes it hard to determine the exact amount of offers that should be
> rescinded in the master.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)