Benjamin Mahler created MESOS-429:
-------------------------------------
Summary: Hadoop MesosScheduler has a deadlock.
Key: MESOS-429
URL: https://issues.apache.org/jira/browse/MESOS-429
Project: Mesos
Issue Type: Bug
Reporter: Benjamin Mahler
Assignee: Benjamin Mahler
Priority: Blocker
This was found with the help of Brenden Matthews.
JobTracker.heartbeat (synchronized) calls MesosScheduler.assignTasks
(synchronized)
MesosScheduler.resourceOffers (synchronized) calls into JobTracker.getJobStatus
(synchronized).
Thread 24558: (state = BLOCKED)
- org.apache.hadoop.mapred.JobTracker.getJobStatus(java.util.Collection,
boolean) @bci=0, line=4592 (Interpreted frame)
- org.apache.hadoop.mapred.JobTracker.jobsToComplete() @bci=11, line=4157
(Interpreted frame)
-
org.apache.hadoop.mapred.MesosScheduler.resourceOffers(org.apache.mesos.SchedulerDriver,
java.util.List) @bci=9, line=273 (Compiled frame)
Thread 24575: (state = BLOCKED)
-
org.apache.hadoop.mapred.MesosScheduler.assignTasks(org.apache.hadoop.mapreduce.server.jobtracker.TaskTracker)
@bci=25, line=219 (Compiled frame)
-
org.apache.hadoop.mapred.JobTracker.heartbeat(org.apache.hadoop.mapred.TaskTrackerStatus,
boolean, boolean, boolean, short) @bci=507, line=2951 (Compiled frame)
The simplest fix for now would be to unsynchronize the Scheduler interface
implementations. As a result, when we have to modify the state of
MesosScheduler inside those methods, we need to do so in a synchronized block.
So long as we don't invoke the JobTracker methods from these synchronized
blocks, we won't have a deadlock. We can clean this up later, if a cleaner
abstraction is needed.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
