This is related to bpo-29988, and I'm happy to move this to there. I made this a separate issue because this is a workaround, not a fix as was being discussed there. Also unlike bpo-29988, this is not restricted to context managers or finally blocks.

TL;DR: Raising exceptions from interrupt handlers (most notably KeyboardInterrupt) can wreak havoc in ways that are impossible to fix. This should be noted in the documentation, with a workaround.

I've attached a few example scripts that cause various strange behavior on Linux when a KeyboardInterrupt is raised at just the right time. There are likely many, many more possible examples:
  - sigint_condition_1.py: Cause a deadlock with threading.Condition
  - sigint_condition_2.py: Cause a double-release and/or notify on unacquired threading.Condition
  - sigint_tempfile.py: Cause NamedTemporaryFiles to not be deleted
  - sigint_zipfile.py: Cause ZipExtFile to corrupt its state

When a user presses Ctrl+C, a KeyboardInterrupt will be raised on the main thread at some later time. This exception may be raised after any bytecode, and most Python code, including the standard library, is not designed to handle exceptions that spring up from nowhere.

As a simple example, consider threading.Condition:

    def __enter__(self):
        return self._lock.__enter__()

The KeyboardInterrupt could be raised just prior to return. In this case, __exit__ will never be called, and the underlying lock will remain acquired. A similar problem occurs if KeyboardInterrupt occurs at the start of __exit__.

This can be mitigated by attempting to catch a KeyboardInterrupt *absolutely everywhere*, but even then, it can't be fixed completely.

    def __enter__(self):
        try:
            # it could happen here, in which case we should not unlock
            ret = self._lock.__enter__()
            # it could happen here, in which case we must unlock
        except KeyboardInterrupt:
	    # it could, in theory, happen again right here
            ...
	    raise
	return ret
        # it could happen here, which is the same problem we had before

This is not restricted to context handlers or try/finally blocks. The zipfile module is a good example of code that is almost certain to enter an inconsistent state if a KeyboardInterrupt is raised while it's doing work:

    class ZipExtFile:
        ...
        def read1(self, n):
            ...
            self._readbuffer = b''
            # what happens if KeyboardInterrupt happens here?
            self._offset = 0
            ...

Due to how widespread this is, it's not worth "fixing". (And honestly, it seems to be a rare problem in practice.) I believe that it would be better to clearly document that KeyboardInterrupt (or any exception propagated from a signal handler) may leave the system
in an inconsistent state. Complex or high reliability applications should avoid catching KeyboardInterrupt as a way of gracefully shutting down, and should prefer registering their own SIGINT handler. They should also avoid raising exceptions from signal handlers at all.