**If** we add a callback API for this, I agree with vstinner's https://bugs.python.org/issue42969#msg402558 commentary on the other callback APIs.  We can do this one similarly and keep it simple. - Why? It's the initial simplicity that provides those running into this problem a way to solve their problem today.  Without getting hung up on the details of the new default behavior.

**That said** the only conclusion I'm coming to for what the safest behavior of Python is in this situation is to hang the threads, effectively as PR 28525 is proposing.

So even if we added the callback API, I'd expect code using Python with C++ where these issues might ever crop up to always register a thread hanging callback.  Which really suggests it is the good default.  So do we even need the callback?

...But lets put the default behavior choice aside for a moment, there's something valuable regardless...

There are **New C APIs to enhance PyGILState_Ensure** proposed in PR 28525 to do with GIL acquisition from threads.  These appear useful as is.  They provide a way for threads to discover that they will never be able to get the GIL and re-enter the Python interpreter.  Rather than today's unexpected behavior of PyGILState_Ensure mercilessly terminating their thread, or given a Callback API whatever effect such a Callback API would have.  That allows code to be written with pre-problem-detection that avoids entering this point of no return state.  That is good for everyone.  **We should add those "GIL-curious" APIs no matter what.**  This bit could become its own PR separate from the other parts.

If we cannot agree that blocking a non-main daemon-or-C/C++ thread forever that is guaranteed to not acquire the GIL because the Python interpreter is going away is the right default behavior instead of blindly killing it under the unsupportable assumption that it has nothing to clean up: We can proceed with such a callback API. I'm having a hard time imagining any other behavior that makes sense, so I'd expect lots of Python extension interface code to start carrying a copy of an implementation of a hang callback and sporting a Py_SetThreadExitCallback(xxx) call in their module Init function. (especially if pybind11 finds needs to generate this as boilerplate)

I think the pthread_exit() call added in issue1856's https://hg.python.org/cpython/rev/c892b0321d23 were well intentioned, but not aware of the full unsupportable ramifications of that API call. That the attempt at releasing a backport of the pthread_exit call to 2.7.8 broke an application (that was admittedly doing something unwise) and had to be reverted as 2.7.9 was a sign for us to revisit this in 3.x releases as well.  Which we're finally doing here.