[
https://issues.apache.org/jira/browse/YARN-9088?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16819447#comment-16819447
]
Karthik Palaniappan commented on YARN-9088:
-------------------------------------------
+1. I think we should consider rolling back YARN-6467 instead of fixing it.
I believe the original behavior was correct – metrics for the root queue should
include metrics for all child queues and partitions. So AllocatedMB /
AvailableMB, for example, give you a global view of cluster utilization. If
YARN-6492 ever gets submitted, then we'll get per-partition metrics too. But I
think YARN-6467 is the worst of both worlds – you don't get per partition
metrics, and you don't get a global view of the cluster.
A lot of cloud providers use cluster-level YARN metrics for autoscaling, and
YARN-6467 breaks autoscaling.
Side note: YARN-6467 was a breaking change with no documentation / release
note. So rolling it back (another breaking change) should be fine. I'll attach
a patch, as long as the rollback is straightforward.
> Non-exclusive labels break QueueMetrics
> ---------------------------------------
>
> Key: YARN-9088
> URL: https://issues.apache.org/jira/browse/YARN-9088
> Project: Hadoop YARN
> Issue Type: Bug
> Components: capacity scheduler, resourcemanager
> Affects Versions: 2.8.5
> Reporter: Brandon Scheller
> Priority: Major
> Labels: metrics, nodelabel
>
> QueueMetrics are broken (random/negative values) when non-exclusive labels
> are being used and unlabeled containers run on labeled nodes.
> This is caused by the change in the patch here:
> https://issues.apache.org/jira/browse/YARN-6467
> It assumes that a container's label will be the same as the node's label that
> it is running on.
> If you look within the patch, sometimes metrics are updated using the
> request.getNodeLabelExpression(). And sometimes they are updated using
> node.getPartition().
> This means that in the case where the node is labeled while the container
> request isn't, these metrics only get updated when referring to the default
> queue. This stops metrics from balancing out and results in incorrect and
> negative values in QueueMetrics.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
