Eric Payne commented on YARN-3275:

This situation can happen under the following conditions:

- All of the resources in the cluster are being used
- {{QueueA}} is preemptable and over its absolute capacity (AKA guaranteed 
- {{QueueB}} is not preemptable, over its absolute capacity, also over its 
absolute max capacity (which can happen), and asking for more resources

In the above scenario, {{ProportionalCapacityPreemptionPolicy}} will subtract 
{{QueuB}}'s ideal assigned value from its absolute max capacity value and get a 
negative number, which will adjust it's ideally assigned resources downwards by 
that amount, which will result in that amount getting preempted.

Regardless of the reason, if a queue is marked as unpreemptable, resources 
should never be preempted from that queue.

> Preemption happening on non-preemptable queues
> ----------------------------------------------
>                 Key: YARN-3275
>                 URL: https://issues.apache.org/jira/browse/YARN-3275
>             Project: Hadoop YARN
>          Issue Type: Bug
>    Affects Versions: 2.7.0
>            Reporter: Eric Payne
>            Assignee: Eric Payne
> YARN-2056 introduced the ability to turn preemption on and off at the queue 
> level. In cases where a queue goes over its absolute max capacity (YARN:3243, 
> for example), containers can be preempted from that queue, even though the 
> queue is marked as non-preemptable.
> We are using this feature in large, busy clusters and seeing this behavior.

This message was sent by Atlassian JIRA

Reply via email to