When starting an interactive session, a fake ping job is submitted to determine the state of the repl driver. However, if yarn is out of resources and the repl application is stuck in Accepted state, the ping job will fail and session will go into dead state. However, the yarn application is not killed. As a result, when Yarn got resources back this application will start running and would be leaked.
Need to make sure the application is killed in Yarn before transition to dead in this case.