[SPARK-6492] SparkContext.stop() can deadlock when DAGSchedulerEventProcessLoop dies

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.3.0, 1.4.0
• Fix Version/s: 1.4.0
• Component/s: Spark Core
• Assignee: Ilya Ganelin
• Reporter: Josh Rosen
• Created: 24/Mar/15 07:23
• Updated: 03/Apr/15 18:24
• Resolved: 03/Apr/15 18:24

Description

A deadlock can occur when DAGScheduler death causes a SparkContext to be shut down while user code is concurrently racing to stop the SparkContext in a finally block.

For example:

try {
      sc = new SparkContext("local", "test")
      // start running a job that causes the DAGSchedulerEventProcessor to crash
      someRDD.doStuff()
    }
} finally {
  sc.stop() // stop the sparkcontext once the failure in DAGScheduler causes the above job to fail with an exception
}

This leads to a deadlock. The event processor thread tries to lock on the SparkContext.SPARK_CONTEXT_CONSTRUCTOR_LOCK and becomes blocked because the thread that holds that lock is waiting for the event processor thread to join:

"dag-scheduler-event-loop" daemon prio=5 tid=0x00007ffa69456000 nid=0x9403 waiting for monitor entry [0x00000001223ad000]
   java.lang.Thread.State: BLOCKED (on object monitor)
	at org.apache.spark.SparkContext.stop(SparkContext.scala:1398)
	- waiting to lock <0x00000007f5037b08> (a java.lang.Object)
	at org.apache.spark.scheduler.DAGSchedulerEventProcessLoop.onError(DAGScheduler.scala:1412)
	at org.apache.spark.util.EventLoop$$anon$1.run(EventLoop.scala:52)

"pool-1-thread-1-ScalaTest-running-SparkContextSuite" prio=5 tid=0x00007ffa69864800 nid=0x5903 in Object.wait() [0x00000001202dc000]
   java.lang.Thread.State: WAITING (on object monitor)
	at java.lang.Object.wait(Native Method)
	- waiting on <0x00000007f4b28000> (a org.apache.spark.util.EventLoop$$anon$1)
	at java.lang.Thread.join(Thread.java:1281)
	- locked <0x00000007f4b28000> (a org.apache.spark.util.EventLoop$$anon$1)
	at java.lang.Thread.join(Thread.java:1355)
	at org.apache.spark.util.EventLoop.stop(EventLoop.scala:79)
	at org.apache.spark.scheduler.DAGScheduler.stop(DAGScheduler.scala:1352)
	at org.apache.spark.SparkContext.stop(SparkContext.scala:1405)
	- locked <0x00000007f5037b08> (a java.lang.Object)
[...]

Issue Links

• links to: [Github] Pull Request #5277 (ilganeli)