[SPARK-21563] Race condition when serializing TaskDescriptions and adding jars

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.2.0
• Fix Version/s: 2.2.1, 2.3.0
• Component/s: Scheduler, Spark Core
• Assignee: Andrew Ash
• Reporter: Andrew Ash
• Created: 28/Jul/17 20:08
• Updated: 14/Aug/17 14:58
• Resolved: 14/Aug/17 14:58

Description

cc robert3005

I was seeing this exception during some running Spark jobs:

16:16:28.294 [dispatcher-event-loop-14] ERROR org.apache.spark.rpc.netty.Inbox - Ignoring error
java.io.EOFException: null
    at java.io.DataInputStream.readFully(DataInputStream.java:197)
    at java.io.DataInputStream.readUTF(DataInputStream.java:609)
    at java.io.DataInputStream.readUTF(DataInputStream.java:564)
    at org.apache.spark.scheduler.TaskDescription$$anonfun$decode$1.apply(TaskDescription.scala:127)
    at org.apache.spark.scheduler.TaskDescription$$anonfun$decode$1.apply(TaskDescription.scala:126)
    at scala.collection.immutable.Range.foreach(Range.scala:160)
    at org.apache.spark.scheduler.TaskDescription$.decode(TaskDescription.scala:126)
    at org.apache.spark.executor.CoarseGrainedExecutorBackend$$anonfun$receive$1.applyOrElse(CoarseGrainedExecutorBackend.scala:95)
    at org.apache.spark.rpc.netty.Inbox$$anonfun$process$1.apply$mcV$sp(Inbox.scala:117)
    at org.apache.spark.rpc.netty.Inbox.safelyCall(Inbox.scala:205)
    at org.apache.spark.rpc.netty.Inbox.process(Inbox.scala:101)
    at org.apache.spark.rpc.netty.Dispatcher$MessageLoop.run(Dispatcher.scala:213)
    at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
    at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
    at java.lang.Thread.run(Thread.java:748)

After some debugging, we determined that this is due to a race condition in task serde. cc irashid kayousterhout who last touched that code in SPARK-19796

The race is between adding additional jars to the SparkContext and serializing the TaskDescription.

Consider this sequence of events:

• TaskSetManager creates a TaskDescription using a reference to the SparkContext's jars: https://github.com/apache/spark/blob/v2.2.0/core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala#L506
• TaskDescription starts serializing, and begins writing jars: https://github.com/apache/spark/blob/v2.2.0/core/src/main/scala/org/apache/spark/scheduler/TaskDescription.scala#L84
• the size of the jar map is written out: https://github.com/apache/spark/blob/v2.2.0/core/src/main/scala/org/apache/spark/scheduler/TaskDescription.scala#L63
• on another thread: the application adds a jar to the SparkContext's jars list
• then the entries in the jars list are serialized out: https://github.com/apache/spark/blob/v2.2.0/core/src/main/scala/org/apache/spark/scheduler/TaskDescription.scala#L64

The problem now is that the jars list is serialized as having N entries, but actually N+1 entries follow that count!

This causes task deserialization to fail in the executor, with the stacktrace above.

The same issue also likely exists for files, though I haven't observed that and our application does not stress that codepath the same way it did for jar additions.

One fix here is that TaskSetManager could make an immutable copy of the jars list that it passes into the TaskDescription constructor, so that list doesn't change mid-serialization.