[YARN-7239] Possible launch/cleanup race condition in ContainersLauncher

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: nodemanager
• Labels: newbie
• Assignee: Unassigned
• Reporter: Miklos Szegedi
• Created: 21/Sep/17 19:36
• Updated: 27/Oct/22 07:00

Description

ContainersLauncher.handle() submits the launch job and then adds the job into the collection risking that the cleanup will miss it and return. This should be in reversed order in all 3 instances:

        containerLauncher.submit(launch);
        running.put(containerId, launch);

The cleanup code that the above code is racing with:

        ContainerLaunch runningContainer = running.get(containerId);
        if (runningContainer == null) {
          // Container not launched. So nothing needs to be done.
          LOG.info("Container " + containerId + " not running, nothing to signal.");
          return;
        }
...