[SPARK-32571] yarnClient.killApplication(appId) is never called

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.4.0, 3.0.0
• Component/s: Spark Submit, YARN
• Assignee: Unassigned
• Reporter: A Tester
• Created: 07/Aug/20 23:04
• Updated: 10/Aug/20 23:09

Description

Problem Statement: 

When an application is submitted using spark-submit in cluster mode using yarn, the spark application continues to run on the cluster, even if spark-submit itself has been requested to shutdown (Ctrl-C/SIGTERM/etc.)

While there is code inside org.apache.spark.deploy.yarn.Client.scala that would lead you to believe the spark application on the cluster will shut down, this code is not currently reachable.

Example of behavior:

spark-submit ...

<Ctrl-C> or kill -15 <pid>

spark-submit itself dies

job can still be found running on the cluster

 

Expectation:

When spark-submit is in monitoring a yarn app and spark-submit itself is requested to shutdown (SIGTERM, HUP,etc.), it should call yarnClient.killApplication(appId) so that the actual spark application running on the cluster is killed.

 

 

Proposal

There is already a shutdown hook registered which cleans up temp files.  Could this be extended to call yarnClient.killApplication? 

I believe the default behavior should be to request yarn to kill the application, however I can imagine use cases where you may still want it to run.  So facilitate these use cases, an option should be provided to skip this hook.