spark序列化异常和Executor的僵死问题

在Spark上执行hive语句的时候，出现类似于如下的异常：

1. org.apache.spark.SparkDriverExecutionException: Execution error
   
2.     at org.apache.spark.scheduler.DAGScheduler.handleTaskCompletion(DAGScheduler.scala:849)
   
3.     at org.apache.spark.scheduler.DAGSchedulerEventProcessActor$anonfun$receive$2.applyOrElse(DAGScheduler.scala:1231)
   
4.     at akka.actor.ActorCell.receiveMessage(ActorCell.scala:498)
   
5.     at akka.actor.ActorCell.invoke(ActorCell.scala:456)
   
6.     at akka.dispatch.Mailbox.processMailbox(Mailbox.scala:237)
   
7.     at akka.dispatch.Mailbox.run(Mailbox.scala:219)
   
8.     at akka.dispatch.ForkJoinExecutorConfigurator$AkkaForkJoinTask.exec(AbstractDispatcher.scala:386)
   
9.     at scala.concurrent.forkjoin.ForkJoinTask.doExec(ForkJoinTask.java:260)
   
10.     at scala.concurrent.forkjoin.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:1339)
    
11.     at scala.concurrent.forkjoin.ForkJoinPool.runWorker(ForkJoinPool.java:1979)
    
12.     at scala.concurrent.forkjoin.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:107)
    
13. Caused by: java.lang.ClassCastException: scala.collection.mutable.HashSet cannot be cast to scala.collection.mutable.BitSet
    
14.     at org.apache.spark.sql.execution.BroadcastNestedLoopJoin$anonfun$7.apply(joins.scala:336)
    
15.     at org.apache.spark.rdd.RDD$anonfun$19.apply(RDD.scala:813)
    
16.     at org.apache.spark.rdd.RDD$anonfun$19.apply(RDD.scala:810)
    
17.     at org.apache.spark.scheduler.JobWaiter.taskSucceeded(JobWaiter.scala:56)
    
18.     at org.apache.spark.scheduler.DAGScheduler.handleTaskCompletion(DAGScheduler.scala:845)

复制代码

排查其前后的日志，发现大都是序列化的东西：

1. 14/08/13 11:10:01 INFO org.apache.spark.Logging$class.logInfo(Logging.scala:58): Serialized task 8.0:3 as 20849 bytes in 0 ms
   
2. 14/08/13 11:10:01 INFO org.apache.spark.Logging$class.logInfo(Logging.scala:58): Finished TID 813 in 25 ms on sparktest0 (progress: 3/200)

复制代码

而在 spark-default.conf中，事先设置了序列化方式为Kryo：

1. spark.serializer org.apache.spark.serializer.KryoSerializer

复制代码

根据异常信息，可见是HashSet转为BitSet类型转换失败，Kryo把松散的HashSet转换为了紧凑的BitSet，把序列化方式注释掉之后，任务可以正常执行。难道Spark的Kryo序列化做的还不到位？ 此问题需要进一步跟踪。

9 Executor僵死问题

    运行一个Spark任务，发现其运行速度远远慢于执行同样SQL语句的Hive的执行，甚至出现了OOM的错误，最后卡住达几小时！并且Executor进程在疯狂GC。

    截取其一Task的OOM异常信息：

可以看到这是在序列化过程中发生的OOM。根据节点信息，找到对应的Executor进程，观察其Jstack信息：

1. Thread 36169: (state = BLOCKED)
   
2. - java.lang.Long.valueOf(long) @bci=27, line=557 (Compiled frame)
   
3. - com.esotericsoftware.kryo.serializers.DefaultSerializers$LongSerializer.read(com.esotericsoftware.kryo.Kryo, com.esotericsoftware.kryo.io.Input, java.lang.Class) @bci=5, line=113 (Compiled frame)
   
4. - com.esotericsoftware.kryo.serializers.DefaultSerializers$LongSerializer.read(com.esotericsoftware.kryo.Kryo, com.esotericsoftware.kryo.io.Input, java.lang.Class) @bci=4, line=103 (Compiled frame)
   
5. - com.esotericsoftware.kryo.Kryo.readClassAndObject(com.esotericsoftware.kryo.io.Input) @bci=158, line=732 (Compiled frame)
   
6. - com.esotericsoftware.kryo.serializers.DefaultArraySerializers$ObjectArraySerializer.read(com.esotericsoftware.kryo.Kryo, com.esotericsoftware.kryo.io.Input, java.lang.Class) @bci=158, line=338 (Compiled frame)
   
7. - com.esotericsoftware.kryo.serializers.DefaultArraySerializers$ObjectArraySerializer.read(com.esotericsoftware.kryo.Kryo, com.esotericsoftware.kryo.io.Input, java.lang.Class) @bci=4, line=293 (Compiled frame)
   
8. - com.esotericsoftware.kryo.Kryo.readObject(com.esotericsoftware.kryo.io.Input, java.lang.Class, com.esotericsoftware.kryo.Serializer) @bci=136, line=651 (Compiled frame)
   
9. - com.esotericsoftware.kryo.serializers.FieldSerializer$ObjectField.read(com.esotericsoftware.kryo.io.Input, java.lang.Object) @bci=143, line=605 (Compiled frame)
   
10. - com.esotericsoftware.kryo.serializers.FieldSerializer.read(com.esotericsoftware.kryo.Kryo, com.esotericsoftware.kryo.io.Input, java.lang.Class) @bci=44, line=221 (Compiled frame)
    
11. - com.esotericsoftware.kryo.Kryo.readObject(com.esotericsoftware.kryo.io.Input, java.lang.Class, com.esotericsoftware.kryo.Serializer) @bci=136, line=651 (Compiled frame)
    
12. - com.esotericsoftware.kryo.serializers.FieldSerializer$ObjectField.read(com.esotericsoftware.kryo.io.Input, java.lang.Object) @bci=143, line=605 (Compiled frame)
    
13. - com.esotericsoftware.kryo.serializers.FieldSerializer.read(com.esotericsoftware.kryo.Kryo, com.esotericsoftware.kryo.io.Input, java.lang.Class) @bci=44, line=221 (Compiled frame)
    
14. - com.esotericsoftware.kryo.Kryo.readClassAndObject(com.esotericsoftware.kryo.io.Input) @bci=158, line=732 (Compiled frame)
    
15. - org.apache.spark.serializer.KryoDeserializationStream.readObject(scala.reflect.ClassTag) @bci=8, line=118 (Compiled frame)
    
16. - org.apache.spark.serializer.DeserializationStream$anon$1.getNext() @bci=10, line=125 (Compiled frame)
    
17. - org.apache.spark.util.NextIterator.hasNext() @bci=16, line=71 (Compiled frame)
    
18. - org.apache.spark.storage.BlockManager$LazyProxyIterator$1.hasNext() @bci=4, line=1031 (Compiled frame)
    
19. - scala.collection.Iterator$anon$13.hasNext() @bci=4, line=371 (Compiled frame)
    
20. - org.apache.spark.util.CompletionIterator.hasNext() @bci=4, line=30 (Compiled frame)
    
21. - org.apache.spark.InterruptibleIterator.hasNext() @bci=22, line=39 (Compiled frame)
    
22. - scala.collection.Iterator$anon$11.hasNext() @bci=4, line=327 (Compiled frame)
    
23. - org.apache.spark.sql.execution.HashJoin$anonfun$execute$1.apply(scala.collection.Iterator, scala.collection.Iterator) @bci=14, line=77 (Compiled frame)
    
24. - org.apache.spark.sql.execution.HashJoin$anonfun$execute$1.apply(java.lang.Object, java.lang.Object) @bci=9, line=71 (Interpreted frame)
    
25. - org.apache.spark.rdd.ZippedPartitionsRDD2.compute(org.apache.spark.Partition, org.apache.spark.TaskContext) @bci=48, line=87 (Interpreted frame)
    
26. - org.apache.spark.rdd.RDD.computeOrReadCheckpoint(org.apache.spark.Partition, org.apache.spark.TaskContext) @bci=26, line=262 (Interpreted frame)

复制代码

有大量的 BLOCKED线程，继续观察GC信息，发现大量的 FULL GC。

    分析，在插入Hive表的时候，实际上需要写HDFS，在此过程的HashJoin时，伴随着大量的Shuffle写操作，JVM的新生代不断GC，Eden Space写满了就往Survivor Space写，同时超过一定大小的数据会直接写到老生代，当新生代写满了之后，也会把老的数据搞到老生代，如果老生代空间不足了，就触发FULL GC，还是空间不够，那就OOM错误了，此时线程被Blocked，导致整个Executor处理数据的进程被卡住。

当处理大数据的时候，如果 JVM配置不当就容易引起上述问题。解决的方法就是增大Executor的使用内存，合理配置新生代和老生代的大小，可以将老生代的空间适当的调大点