[ 
https://issues.apache.org/jira/browse/IMPALA-2968?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16872546#comment-16872546
 ] 

Andrew Sherman commented on IMPALA-2968:
----------------------------------------

If the head of the queue needs to wait for resources, then this algorithm will 
delay the time when it blocks other smaller queries. 
So one issue is that this only delays the time when this happens. 
It also is hard to reason about the configuration, an administrator would have 
to have some understanding of how often this counter is incremented, which 
might be confusing.
In an ideal world there might be AC policies which control behavior at a more 
descriptive level.

So the suggested algorithm seems OK but the question is whether it brings 
enough benefit for the added complexity

> Improve admission control dequeuing policy
> ------------------------------------------
>
>                 Key: IMPALA-2968
>                 URL: https://issues.apache.org/jira/browse/IMPALA-2968
>             Project: IMPALA
>          Issue Type: Improvement
>          Components: Backend
>    Affects Versions: Impala 2.3.0
>            Reporter: Matthew Jacobs
>            Priority: Minor
>              Labels: admission-control, resource-management
>
> The current behavior only attempts to admit the head of the queue but the 
> head might need resources which are contended (e.g. on a hot node) while a 
> queued request behind the head might be admitted. We should consider a policy 
> which would not block the entire queue but yet is unlikely to starve the head 
> if other requests are continuously admitted from behind it.



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

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org

Reply via email to