Unfortunately, getting accurate state information back from the cluster can be tricky. This is because there are multiple sources of information relating to this - there's the "JobX/TaskY.state.mat" files on disk in your JobStorageLocation. These are created in state pending, the client moves them to queued on submission, and then the worker MATLAB processes set them to be running, and finally finished. There's also the information coming back from querying the underlying scheduling system. These pieces of information can occasionally (and usually transiently) conflict with each other, which leads to spurious states being observed. (It is necessary to query the underlying scheduling system to deal with the case where the worker MATLAB crashes before it gets to set the state file to running or finished.)
If you can, I would recommend using Job.wait as your primary means of waiting for results to become available. (Perhaps with the timeout parameter). This method ought to be more reliable than querying the task State properties directly, as it performs more detailed (and more expensive) checks.
