[ 
https://issues.apache.org/jira/browse/MESOS-9748?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16829024#comment-16829024
 ] 

Benjamin Bannier commented on MESOS-9748:
-----------------------------------------

[~chhsia0], one could also argue that a framework which let's offers time out 
instead of accepting or declining them has already violated the offer protocol. 
I agree that implicit behavior can be confusing, but I believe one could argue 
here from a position of Mesos guarding itself against misbehaving frameworks 
(especially important since we rely on frameworks behaving correctly).

> Consider installing filters if offers time out
> ----------------------------------------------
>
>                 Key: MESOS-9748
>                 URL: https://issues.apache.org/jira/browse/MESOS-9748
>             Project: Mesos
>          Issue Type: Improvement
>          Components: allocation, master
>            Reporter: Benjamin Bannier
>            Priority: Major
>
> We currently do not install any filters if an offer to a framework times out. 
> We should consider recovering with a filter for some duration.
> This can be interesting in cases where e.g., framework offer processing is 
> backlogged so it cannot accept offers in time. If no filter is installed we 
> could end up allocating to that framework again and again (and exclusively if 
> other frameworks already use more resources than what we offer the backlogged 
> framework at a time).
> If we would instead filter the resources from timed out offers for some 
> allocation cycle, we could extert some form of backpressure on the 
> framework's behalf which could allow it to recover. We also would give a 
> chance that other frameworks can actually use the resources.
> The optimal duration to filter the resources for would depend on the 
> (dynamic) number of frameworks interested in offers and might be hard to 
> determine. Filtering for a single allocation cycle might be a good starting 
> point to get some of the benefits without causing too much new behavior.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to