Phil D'Amore created YARN-2726:
----------------------------------
Summary: CapacityScheduler should explicitly log when an
accessible label has no capacity
Key: YARN-2726
URL: https://issues.apache.org/jira/browse/YARN-2726
Project: Hadoop YARN
Issue Type: Improvement
Components: capacityscheduler
Reporter: Phil D'Amore
Priority: Minor
Given:
- Node label defined: test-label
- Two queues defined: a, b
- label accessibility and and capacity defined as follows (properties
abbreviated for readability):
root.a.accessible-node-labels = test-label
root.a.accessible-node-labels.test-label.capacity = 100
If you restart the RM or do a 'rmadmin -refreshQueues' you will get a stack
trace with the following error buried within:
"Illegal capacity of -1.0 for label=test-label in queue=root.b"
This of course occurs because test-label is accessible to b due to inheritance
from the root, and -1 is the UNDEFINED value. To my mind this might not be
obvious to the admin, and the error message which results does not help guide
someone to the source of the issue.
I propose that this situation be updated so that when the capacity on an
accessible label is undefined, it is explicitly called out instead of falling
through to the illegal capacity check. Something like:
{code}
if (capacity == UNDEFINED) {
throw new IllegalArgumentException("Configuration issue: " + " label=" +
label + " is accessible from queue=" + queue + " but has no capacity set.");
}
{code}
I'll leave it to better judgement than mine as to whether I'm throwing the
appropriate exception there. I think this check should be added to both
getNodeLabelCapacities and getMaximumNodeLabelCapacities in
CapacitySchedulerConfiguration.java.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
