如何可靠地终止使用 spark-submit 提交的 spark 应用程序

Question

I'm seeing a strange problem. I have a spark cluster in standalone mode. I submit spark jobs from a remote node as follows from the terminal

$> spark-submit --master spark://10.1.40.18:7077  --class com.test.Ping spark-jobs.jar

when the app is running , when I press ctrl-C on the console terminal, then the process is killed and so is the app in the spark master UI. When I go to spark master ui, i see that this app is in state Killed under Completed applications, which is what I expected to see.

Now, I created a shell script as follows to do the same

#!/bin/bash
spark-submit --master spark://10.1.40.18:7077  --class com.test.Ping spark-jobs.jar &
echo $! > my.pid

When I execute the shell script from terminal, as follows

$> bash myscript.sh

The application is submitted correctly to spark master and I can see it as one of the running apps in the spark master UI. But when I kill the process in my terminal as follows

$> ps kill $(cat my.pid)

I see that the process is killed on my machine but the spark application is still running in spark master! It doesn't get killed.

I noticed one more thing that, when I launch the spark job via shell script and kill the application from spark master UI by clicking on "kill" next to the running application, it gets killed in spark ui but I still see the process running in my machine.

In both cases, I would expect the remote spark app to be killed and my local process to be killed.

Why is this happening? and how can I kill a spark app from the terminal launced via shell script wo going to the spark master UI?

I want to launch the spark app via script and log the pid so i can monitor it remotely

thanks for the help

Answer 1

I solved the first issue by adding a shutdown hook in my code. The shutdown hook get call when you exit your script (ctrl-C , kill … but nor kill -9)

val shutdownHook = scala.sys.addShutdownHook {
try {

        sparkContext.stop()
//Make sure to kill any other threads or thread pool you may be running 
      }
      catch {
        case e: Exception =>
          {
            ...

          }
      }

    }

For the other issue , kill from the UI. I also had the issue. This was caused by a thread pool that I use.

So I surrounded my code with try/finally block to guarantee that the thread pool was shutdown when spark stopped

I hope this helps

Answer 2

使用spark杀死任务的方法是：

park-submit --kill [submission ID] --master [spark://...]

Answer 3

This is the approach I follow when I want to kill a specific SPARK job that is running on the cluster mode and with the new version of the application I want to start it again, so handling this programmatically is the best way to do it.

I follow two shell scripts to stop and start .

the start.sh is similar to what we all know spark-submit script

Another one is stop.sh and code is below.

I invoke stop.sh first and then I invoke start.sh (via CI/CD), that is up to you how you automate end to end deployment.

This piece of code will kill a particular job (if running) given its name as described in the spark-submit "name" parameter.

Also to be on a safer side I propose trimming of any leading or trailing whitespace. Execute these in the same cluster where the SPARK jobs are running.

#!/bin/bash
echo "stop.sh invoked"
export APPLICATION_NAME=my-spark-job-name
echo "Killing $APPLICATION_NAME"
APPLICATION_ID=$(yarn application --appStates RUNNING --list 2>/dev/null | awk "{ if (\$2 == \"$APPLICATION_NAME\") print \$APPLICATION_NAME }")
if ["$APPLICATION_ID" = ""]
then
        echo "$APPLICATION_NAME not running."
else
        APPLICATION_ID="${APPLICATION_ID#"${APPLICATION_ID%%[![:space:]]*}"}"
        APPLICATION_ID="${APPLICATION_ID%"${APPLICATION_ID##*[![:space:]]}"}"
        yarn application --kill $APPLICATION_ID 2>/dev/null
        echo $APPLICATION_NAME " Successfully Killed!"
fi