I am using Spark 2.0.0 with Hadoop 2.7 and use the yarn-cluster mode. Every time, I get the following error:
17/01/04 11:18:04 INFO spark.SparkContext: Successfully stopped SparkContext 17/01/04 11:18:04 INFO yarn.ApplicationMaster: Final app status: FAILED, exitCode: 16, (reason: Shutdown hook called before final status was reported.) 17/01/04 11:18:04 INFO util.ShutdownHookManager: Shutdown hook called 17/01/04 11:18:04 INFO util.ShutdownHookManager: Deleting directory /tmp/hadoop-hduser/nm-local-dir/usercache/harry/appcache/application_1475261544699_0833/spark-42e40ac3-279f-4c3f-ab27-9999d20069b8 17/01/04 11:18:04 INFO spark.SparkContext: SparkContext already stopped.
However, I do get the correct printed output. The same code works fine in Spark 1.4.0-Hadoop 2.4.0 where I do not see any exit codes.
This issue .sparkStaging not cleaned if application exited incorrectly https://issues.apache.org/jira/browse/SPARK-17340 started after Spark 1.4 (Affects Version/s: 1.5.2, 1.6.1, 2.0.0)
The issue is: When running Spark (yarn,cluster mode) and killing application .sparkStaging is not cleaned.
When this issue happened exitCode 16 raised in Spark 2.0.X
ERROR ApplicationMaster: RECEIVED SIGNAL TERM INFO ApplicationMaster: Final app status: FAILED, exitCode: 16, (reason: Shutdown hook called before final status was reported.)
Is it possible that in your code, something is killing the application? If so - it shouldn't be seen in Spark 1.4, but should be seen in Spark 2.0.0
Please search your code for "exit" (as if you have such in your code, the error won't be shown in Spark 1.4, but will in Spark 2.0.0)
Seems like excess memory is used by JVM, try adding the property
yarn.nodemanager.vmem-check-enabled to false in your yarn-site.xml
Reference
