[ 
https://issues.apache.org/jira/browse/MESOS-4102?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15065909#comment-15065909
 ] 

Alexander Rukletsov commented on MESOS-4102:
--------------------------------------------

I see two approaches how we can mitigate current behaviour. One is to persist 
laid aside resources for quota between allocations, so that when a new agent 
joins, we can check whether we can allocate its resources to non-quota'ed 
frameworks. This approach however prevents quota'ed frameworks to "see" the new 
agent and therefore feels a bit wrong, because quota should go first and fair 
share second. The other approach is whenever allocation is triggered, add all 
agents with laid aside resources (for example, if we have laid aside resource 
on {{agent3}} and {{agent6}}, when a new {{agent20}} joins the cluster, we call 
{{allocate()}} not only for {{agent20}}, but for all three: {{agent3}}, 
{{agent6}}, and {{agent20}}). This may increase the time we spend in 
{{allocate()}} but seems to be more fair than the first approach.

> Quota doesn't allocate resources on slave joining
> -------------------------------------------------
>
>                 Key: MESOS-4102
>                 URL: https://issues.apache.org/jira/browse/MESOS-4102
>             Project: Mesos
>          Issue Type: Bug
>          Components: allocation
>            Reporter: Neil Conway
>            Assignee: Alexander Rukletsov
>              Labels: mesosphere, quota
>         Attachments: quota_absent_framework_test-1.patch
>
>
> See attached patch. {{framework1}} is not allocated any resources, despite 
> the fact that the resources on {{agent2}} can safely be allocated to it 
> without risk of violating {{quota1}}. If I understand the intended quota 
> behavior correctly, this doesn't seem intended.
> Note that if the framework is added _after_ the slaves are added, the 
> resources on {{agent2}} are allocated to {{framework1}}.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to