[
https://issues.apache.org/jira/browse/YUNIKORN-1641?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17705307#comment-17705307
]
Peter Bacsko edited comment on YUNIKORN-1641 at 3/27/23 11:34 AM:
------------------------------------------------------------------
That looks different to me. Those checks tell whether a given user/group have
zero tracked resources. We need to remove a tracker from the hierarchy whenever
the tracked resource becomes zero. To me this looks like a good approach inside
{{decreaseTrackedResource()}}:
{noformat}
func (qt *QueueTracker) decreaseTrackedResource(queuePath string, applicationID
string, usage *resources.Resource, removeApp bool) error {
...
for childQueue, tracker := range qt.childQueueTrackers {
if resources.IsZero(tracker.resourceUsage) {
log.Logger().Info("Removing child tracker, resource
usage is zero",
zap.String("this queue", qt.queueName),
zap.String("child queue", childQueue))
delete(qt.childQueueTrackers, childQueue)
}
}
if len(qt.childQueueTrackers) == 0 &&
!resources.IsZero(qt.resourceUsage) {
// invariant violated
return fmt.Errorf("BUG: no child trackers left, but the
resource usage is not zero - tracker %s resource %v",
qt.queueName, qt.resourceUsage)
}
return nil
{noformat}
So every tracker checks its own immediate child trackers.
was (Author: pbacsko):
That looks different to me. Those checks tell whether a given user/group have
zero tracked resources. We need to remove a tracker from the hierarchy whenever
the tracked resource becomes zero. To me this looks like a good approach inside
{{decreaseTrackedResource()}}:
{noformat}
func (qt *QueueTracker) decreaseTrackedResource(queuePath string, applicationID
string, usage *resources.Resource, removeApp bool) error {
...
for childQueue, tracker := range qt.childQueueTrackers {
if resources.IsZero(tracker.resourceUsage) {
log.Logger().Info("Removing tracker, resource usage is
zero",
zap.String("this queue", qt.queueName),
zap.String("child queue", childQueue))
delete(qt.childQueueTrackers, childQueue)
}
}
if len(qt.childQueueTrackers) == 0 &&
!resources.IsZero(qt.resourceUsage) {
// invariant violated
return fmt.Errorf("BUG: no child trackers left, but the
resource usage is not zero - tracker %s resource %v",
qt.queueName, qt.resourceUsage)
}
return nil
{noformat}
So every tracker checks its own immediate child trackers.
> Queue tracker object is not cleaned up when a queue is deleted
> --------------------------------------------------------------
>
> Key: YUNIKORN-1641
> URL: https://issues.apache.org/jira/browse/YUNIKORN-1641
> Project: Apache YuniKorn
> Issue Type: Sub-task
> Components: core - scheduler
> Reporter: Peter Bacsko
> Assignee: Qi Zhu
> Priority: Major
> Labels: pull-request-available
>
> Currently, every queue has its own {{QueueTracker}} object which is managed
> by {{{}ugm.Manager{}}}.
> But when the queue is deleted either by a config change or automatically by
> the partition manager, the tracker object stays in memory, causing a memory
> leak.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
