Spark Kill Running Application

118
votes

I have a running Spark application where it occupies all the cores where my other applications won't be allocated any resource.

I did some quick research and people suggested using YARN kill or /bin/spark-class to kill the command. However, I am using CDH version and /bin/spark-class doesn't even exist at all, YARN kill application doesn't work either.

enter image description here

Can anyone with me with this?

5
apache-sparkhadoop-yarnpyspark
if you are in a test env : ps aux | grep spark -> get the pid of spark and kill it from command line – eliasah
@eliasah "test env", to me the job is already distributed.. – B.Mr.W.
you want to kill a job in production???? – eliasah
@eliasah Yeah... a job in production got hung due the failure in one host. – B.Mr.W.

5 Answers

249
votes
  • copy paste the application Id from the spark scheduler, for instance application_1428487296152_25597
  • connect to the server that have launch the job
  • yarn application -kill application_1428487296152_25597
10
votes

It may be time consuming to get all the application Ids from YARN and kill them one by one. You can use a Bash for loop to accomplish this repetitive task quickly and more efficiently as shown below:

Kill all applications on YARN which are in ACCEPTED state:

for x in $(yarn application -list -appStates ACCEPTED | awk 'NR > 2 { print $1 }'); do yarn application -kill $x; done

Kill all applications on YARN which are in RUNNING state:

for x in $(yarn application -list -appStates RUNNING | awk 'NR > 2 { print $1 }'); do yarn application -kill $x; done

7
votes

First use:

yarn application -list

Note down the application id Then to kill use:

yarn application -kill application_id
1
votes

This might not be an ethical and preferred solution but it helps in environments where you can't access the console to kill the job using yarn application command.

Steps are

Go to application master page of spark job. Click on the jobs section. Click on the active job's active stage. You will see "kill" button right next to the active stage.

This works if the succeeding stages are dependent on the currently running stage. Though it marks job as " Killed By User"