[ 
https://issues.apache.org/jira/browse/SLIDER-1025?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15061911#comment-15061911
 ] 

Steve Loughran commented on SLIDER-1025:
----------------------------------------

Now, what's going to happen if two requests get issued for the same 
role/priority with different resource requirements.

if the largest one comes in first, there's a risk the smaller one will get in, 
because it will fit; the larger one won't get satisfied and when the smaller 
one comes in it will be ignored.

question is: does that matter? Because container sizes are only going to change 
when a flex comes in. I think that check is primarily a safety check. Really we 
should be reacting to a flex by checking to see if the requirements have 
changed —and if so, killing all outstanding requests.

so...
+1 for this patch; I'll merge it in. 

> Outstanding container request is not removed from open list even it is 
> already been allocated
> ---------------------------------------------------------------------------------------------
>
>                 Key: SLIDER-1025
>                 URL: https://issues.apache.org/jira/browse/SLIDER-1025
>             Project: Slider
>          Issue Type: Bug
>          Components: core
>    Affects Versions: Slider 0.80, Slider 0.81
>            Reporter: Weiwei Yang
>            Priority: Critical
>         Attachments: SLIDER-1025.001.patch
>
>
> When there is one or more entries in open requests list, and a container is 
> allocated for an outstanding request, it cannot be removed from openRequests 
> list when the resource allocated is *not matched* with the resource 
> requested.  
> {code:java}
> public synchronized boolean resourceRequirementsMatch(Resource resource) {
>     return issuedRequest != null && 
> issuedRequest.getCapability().equals(resource);
>   }
> {code}
> The problem here is RM doesn't always allocate *matched* resource for a 
> request, it normalized the resource request with multiply of minimal 
> allocation size, so instead of *matched*, it could also be *fits* the 
> resource request.



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

Reply via email to