[
https://issues.apache.org/jira/browse/YARN-3434?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14487802#comment-14487802
]
Wangda Tan commented on YARN-3434:
----------------------------------
[~tgraves], you're right. But I'm wondering why this could happen:
When continousReservation enabled, it will do check in assignContainer:
{code}
if (reservationsContinueLooking && rmContainer == null) {
// we could possibly ignoring parent queue capacity limits when
// reservationsContinueLooking is set.
// If we're trying to reserve a container here, not container will be
// unreserved for reserving the new one. Check limits again before
// reserve the new container
if (!checkLimitsToReserve(clusterResource,
application, capability)) {
return Resources.none();
}
}
{code}
When continousReservation disabled, assignContainers will ensure user-limit
will not be violated.
My point is, *user-limit and queue max capacity are all checked before reserve
new container*. And allocation from reserved container will unreserve before
continue. So I think in your case,
https://issues.apache.org/jira/browse/YARN-3434?focusedCommentId=14485834&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14485834:
job-2 cannot reserve 25 * 12 GB containers. Did I miss anything?
And I've a question about continous reservation checking behavior, may or may
not related to this issue: Now it will try to unreserve all containers under a
user, but actually it will only unreserve at most one container to allocate a
new container. Do you think is it fine to change the logic to be:
When (continousReservation-enabled) && (user.usage + required -
min(max-allocation, user.total-reserved) <=user.limit), assignContainers will
continue. This will prevent doing impossible allocation when user reserved lots
of containers. (As same as queue reservation checking).
> Interaction between reservations and userlimit can result in significant ULF
> violation
> --------------------------------------------------------------------------------------
>
> Key: YARN-3434
> URL: https://issues.apache.org/jira/browse/YARN-3434
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacityscheduler
> Affects Versions: 2.6.0
> Reporter: Thomas Graves
> Assignee: Thomas Graves
> Attachments: YARN-3434.patch
>
>
> ULF was set to 1.0
> User was able to consume 1.4X queue capacity.
> It looks like when this application launched, it reserved about 1000
> containers, each 8G each, within about 5 seconds. I think this allowed the
> logic in assignToUser() to allow the userlimit to be surpassed.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
