06-21-2017 07:54 AM - edited 06-21-2017 11:42 AM
I'm running CDH 5.11.1 with the FairScheduler and preemption enabled.
When the cluster is fully utilized (no free vcores) the resource manager will often end up an infinite loop that looks like this.
2017-06-21 14:09:04,752 INFO org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FSSchedulerNode: Updated reserved container container_1498049738482_0045_01_80455566 on node host: <HOSTNAME>:33692 #containers=40 available=96018 used=142208 for application application_1498049738482_0045 2017-06-21 14:09:04,752 INFO org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FSAppAttempt: Making reservation: node=<HOSTNAME> app_id=application_1498049738482_0045
And by infinite loop, I mean it tries to reserve a container on the same node for the same application 30 times PER MILLISECOND until I restart the RM. This looks a whole lot like YARN-4477 (https://issues.apache.org/jira/browse/YARN-4477), but that fix was backported to CDH 5.11.1
Anyone else seen this? Thanks!
07-18-2017 02:42 PM
07-18-2017 03:24 PM
Never saw negative memory, but that particular cluster was vcore rather than memory constrained. We ended up switching to Capacity Scheduler pretty easily with good results, fwiw.
07-20-2017 07:12 PM
We also encountered this error after upgrading to CDH 5.11.1 and had to roll-back consequently.
Does this bug mean that CDH 5.11.1 is an inherently broken release?
A statement from Cloudera would be great.