[
https://issues.apache.org/jira/browse/MESOS-8950?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16738583#comment-16738583
]
Benjamin Mahler commented on MESOS-8950:
----------------------------------------
Probably we're not going to do anything about this issue in the current model.
The long term "solution" to this would be moving to a "shared state" model
where schedulers see the cluster state and perform operations.
> Framework operations can make resources unallocatable
> -----------------------------------------------------
>
> Key: MESOS-8950
> URL: https://issues.apache.org/jira/browse/MESOS-8950
> Project: Mesos
> Issue Type: Bug
> Components: allocation, master
> Reporter: Benjamin Bannier
> Priority: Minor
>
> The allocator does not offer {{cpus}} or {{mem}} resources smaller than
> certain, fixed sizes. For framework operations, we do not enforce the same
> minimum size constraints which can lead the resources becoming unavailable
> for any future allocations. This behavior seems most pronounced when a
> framework can register in many roles.
> Example:
> * A single multirole framework which can register in any role, e.g., in a
> certain role subhierarchy.
> * Single agent with {{cpus:1.5*MIN_CPUS}} and {{mem:1.5*MIN_MEM}}.
> * Framework is offered all resources and performs a {{RESERVE}} on
> {{cpus:0.5*MIN_CPUS}}. It then changes its role.
> * Same framework behavior in next two offer cycles. All {{cpus}} are then
> reserved for different roles in unallocatable amounts.
> * Last offer will be just for {{mem:1.5*MIN_MEM}}, framework reserves 0.6 of
> these to another role. This fragements the {{mem}} resources as well.
> * No allocatable resources left in cluster.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
