为什么集群模式下的YARN上的Spark会因“线程中的异常”驱动程序“java.lang.NullPointerException”而失败？

Question

I'm using emr-5.4.0 with Spark 2.1.0. I understand what NullPointerException is, this question is about why that was thrown in this particular case.

Cannot really figure out why I got NullPointerException in the driver thread.

I got this weird job failing with this error:

18/03/29 20:07:52 INFO ApplicationMaster: Starting the user application in a separate Thread
18/03/29 20:07:52 INFO ApplicationMaster: Waiting for spark context initialization...
Exception in thread "Driver" java.lang.NullPointerException
    at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
    at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
    at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
    at java.lang.reflect.Method.invoke(Method.java:498)
    at org.apache.spark.deploy.yarn.ApplicationMaster$$anon$2.run(ApplicationMaster.scala:637)
18/03/29 20:07:52 ERROR ApplicationMaster: Uncaught exception:
java.lang.IllegalStateException: SparkContext is null but app is still running!
    at org.apache.spark.deploy.yarn.ApplicationMaster.runDriver(ApplicationMaster.scala:415)
    at org.apache.spark.deploy.yarn.ApplicationMaster.run(ApplicationMaster.scala:254)
    at org.apache.spark.deploy.yarn.ApplicationMaster$$anonfun$main$1.apply$mcV$sp(ApplicationMaster.scala:766)
    at org.apache.spark.deploy.SparkHadoopUtil$$anon$1.run(SparkHadoopUtil.scala:67)
    at org.apache.spark.deploy.SparkHadoopUtil$$anon$1.run(SparkHadoopUtil.scala:66)
    at java.security.AccessController.doPrivileged(Native Method)
    at javax.security.auth.Subject.doAs(Subject.java:422)
    at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1698)
    at org.apache.spark.deploy.SparkHadoopUtil.runAsSparkUser(SparkHadoopUtil.scala:66)
    at org.apache.spark.deploy.yarn.ApplicationMaster$.main(ApplicationMaster.scala:764)
    at org.apache.spark.deploy.yarn.ApplicationMaster.main(ApplicationMaster.scala)
18/03/29 20:07:52 INFO ApplicationMaster: Final app status: FAILED, exitCode: 10, (reason: Uncaught exception: java.lang.IllegalStateException: SparkContext is null but app is still running!)
18/03/29 20:07:52 INFO ApplicationMaster: Unregistering ApplicationMaster with FAILED (diag message: Uncaught exception: java.lang.IllegalStateException: SparkContext is null but app is still running!)
18/03/29 20:07:52 INFO ApplicationMaster: Deleting staging directory hdfs://<ip-address>.ec2.internal:8020/user/hadoop/.sparkStaging/application_1522348295743_0010
18/03/29 20:07:52 INFO ShutdownHookManager: Shutdown hook called
End of LogType:stderr

I submitted this job as this:

spark-submit --deploy-mode cluster --master yarn --num-executors 40 --executor-cores 16 --executor-memory 100g --driver-cores 8 --driver-memory 100g --class <package.class_name> --jars <s3://s3_path/some_lib.jar> <s3://s3_path/class.jar>

And my class looks like this:

class MyClass {

  def main(args: Array[String]): Unit = {
    val c = new MyClass()
    c.process()
  }

  def process(): Unit = {
    val sparkConf = new SparkConf().setAppName("my-test")
    val sparkSession: SparkSession = SparkSession.builder().config(sparkConf).getOrCreate()
    import sparkSession.implicits._
    ....
  }

  ...
}

Answer 1

Change class MyClass to object MyClass and you're done.

While we're at it, I'd also change class MyClass to object MyClass extends App and remove def main(args: Array[String]): Unit (as given by extends App ).

I've reported an improvement for Spark 2.3.0 - [SPARK-23830] Spark on YARN in cluster deploy mode fail with NullPointerException when a Spark application is a Scala class not object - to have it reported nicely to an end user.

---

Digging deeper into how Spark on YARN works, the following message is when the ApplicationMaster of a Spark application starts the driver (you used --deploy-mode cluster --master yarn with spark-submit ).

ApplicationMaster: Starting the user application in a separate Thread

Right after the INFO message you should see another:

ApplicationMaster: Waiting for spark context initialization...

This is part of the driver initialization when the ApplicationMaster runs .

The reason for the exception Exception in thread "Driver" java.lang.NullPointerException is due to the following code :

val mainMethod = userClassLoader.loadClass(args.userClass)
  .getMethod("main", classOf[Array[String]])

My understanding is that mainMethod is null at this point so the following line (where mainMethod is null ) "triggers" NullPointerException :

mainMethod.invoke(null, userArgs.toArray)

The thread is indeed called Driver (as in Exception in thread "Driver" java.lang.NullPointerException ) as set in this line :

userThread.setContextClassLoader(userClassLoader)
userThread.setName("Driver")
userThread.start()

The line numbers differ since I used Spark 2.3.0 to reference the lines while you use emr-5.4.0 with Spark 2.1.0.