Major Description
changeInstances() in it's core creates sequential znode that TwillAppMaster processes in order and until current;y processed request is not satisfied rest of the changeInstances() requests will be sitting not processed.
The problem comes up if user made unreasonable changeInstances() request that will be never satisfied with current YARN capacity. User could have mistyped something.
I suggest to at least put a limit on waiting time for resources to check if there are more requests pending in line to abandon current one and proceed to next.