Tao Yang created YARN-11785:
-------------------------------
Summary: Race condition in QueueMetrics due to non-thread-safe
HashMap causes MetricsException
Key: YARN-11785
URL: https://issues.apache.org/jira/browse/YARN-11785
Project: Hadoop YARN
Issue Type: Bug
Components: capacityscheduler
Affects Versions: 3.4.1, 3.3.6, 3.2.4
Reporter: Tao Yang
Assignee: Tao Yang
In high-concurrency scenarios, the async-scheduling thread may fail to retrieve
an existing PartitionQueueMetric from QueueMetrics#QUEUE_METRICS (returns
null), then attempt to re-register the same queue name. This triggers a
MetricsException ("Duplicate metric name") and causes the async-scheduling
thread to exit unexpectedly. Below is the error stack trace:
{code:java}
ERROR org.apache.hadoop.yarn.server.resourcemanager.ResourceManager: Received
RMFatalEvent of type CRITICAL_THREAD_CRASH, caused by a critical thread,
Thread-12, that exited unexpectedly:
org.apache.hadoop.metrics2.MetricsException: Metrics source
PartitionQueueMetrics,partition=,q0=root,q1=xxx already exists!
at
org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.newSourceName(DefaultMetricsSystem.java:152)
at
org.apache.hadoop.metrics2.lib.DefaultMetricsSystem.sourceName(DefaultMetricsSystem.java:125)
at
org.apache.hadoop.metrics2.impl.MetricsSystemImpl.register(MetricsSystemImpl.java:229)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.QueueMetrics.getPartitionQueueMetrics(QueueMetrics.java:286)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.QueueMetrics.setAvailableResourcesToUser(QueueMetrics.java:529)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.LeafQueue.computeUserLimitAndSetHeadroom(LeafQueue.java:1490)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.LeafQueue.assignContainers(LeafQueue.java:1146)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.ParentQueue.assignContainersToChildQueues(ParentQueue.java:803)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.ParentQueue.assignContainers(ParentQueue.java:623)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.ParentQueue.assignContainersToChildQueues(ParentQueue.java:803)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.ParentQueue.assignContainers(ParentQueue.java:623)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.allocateOrReserveNewContainers(CapacityScheduler.java:1697)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.allocateContainerOnSingleNode(CapacityScheduler.java:1632)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.allocateContainersToNode(CapacityScheduler.java:1787)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.allocateContainersToNode(CapacityScheduler.java:1536)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler.schedule(CapacityScheduler.java:606)
at
org.apache.hadoop.yarn.server.resourcemanager.scheduler.capacity.CapacityScheduler$AsyncScheduleThread.run(CapacityScheduler.java:653)
{code}
The root cause is that QUEUE_METRICS field is implemented with HashMap, which
is not thread-safe. Concurrent put and get operations can lead to visibility
issue. This issue can be fixed by ensuring thread-safe access via
ConcurrentHashMap for QUEUE_METRICS field.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
