Spark on Yarn handle AM being told command from RM

When RM throws ApplicationAttemptNotFoundException for allocate
invocation, making the ApplicationMaster to finish immediately without any
retries.

Author: Devaraj K

yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala

@@ -28,6 +28,7 @@ import org.apache.hadoop.fs.{FileSystem, Path}
 import org.apache.hadoop.yarn.api._
 import org.apache.hadoop.yarn.api.records._
 import org.apache.hadoop.yarn.conf.YarnConfiguration
+import org.apache.hadoop.yarn.exceptions.ApplicationAttemptNotFoundException
 
 import org.apache.spark.rpc._
 import org.apache.spark.{Logging, SecurityManager, SparkConf, SparkContext, SparkEnv,
@@ -370,6 +371,12 @@ private[spark] class ApplicationMaster(
             failureCount = 0
           } catch {
             case i: InterruptedException =>
+            case a: ApplicationAttemptNotFoundException => {
+              val message = "ApplicationAttemptNotFoundException was thrown from Reporter thread.";
+              logError(message, a)
+              finish(FinalApplicationStatus.FAILED, ApplicationMaster.EXIT_REPORTER_FAILURE,
+                message)
+            }
             case e: Throwable => {
               failureCount += 1
               if (!NonFatal(e) || failureCount >= reporterMaxFailures) {