Issue 24900: Raising an exception that cannot be unpickled causes hang in ProcessPoolExecutor

When raising an exception that is unpicklable in a worker process. This used to be the case for all exceptions that have a non-trivial constructor but this was fixed by http://bugs.python.org/issue1692335.

I now have the concrete problem with cx_Oracle in which the error type cx_Oracle._Error doesn't define tp_new which results in a

    UnpicklingError: NEWOBJ class argument has NULL tp_new

I have already reported this fact at https://bitbucket.org/anthony_tuininga/cx_oracle/issues/23.

However, I think that the worker thread should either be able to recover from this (which should be quite easy) or, at the very least, fail cleanly. Currently, when this condition is triggered the worker processes are not killed alongside the parent.

There is an existing report and a patch in the backports project: https://github.com/agronholm/pythonfutures/issues/30#issuecomment-132576545

Actually I think that this may consist of multiple issues:

1. http://bugs.python.org/issue8323 (closed as wont-fix): multiprocessing.Queue should throw a sensible error if a Queue element can't be unpickled, however the user managed to create that
2. The worker process should prevent unpicklable ResultItems to be put on the Queue
3. The parent thread in ProcessPoolExecutor should handle the exceptions by setting its broken flag and at least trying to cleanly shut down

The patch for issue22995 disallows default pickling objects with tp_new == NULL.

That would touch 2, but as this is happening in the exception handling already that would need to be catched and handled by itself. However, one could then do without a "pickle -> unpickle" dance.