Race condition in ClusterResourceManager

When the ClusterResourcedManager receives a notification that a container is started, it moves the container from the "pending queue" to its "running queue".
In the meanwhile, it's possible for another thread to remove the mapping for the key. Here's an example:

NMCallbackThread-1:pendingYarnContainers.remove(key);

NMCallbackThread-2:

for (String key : pendingYarnContainers.keySet())

{ yarnContainer = pendingYarnContainers.get(key); <-- could be null depending on whether the removal happened before it. }