FWIW, that atexit callbacks were not called for sub interpreters ever was a pain point for mod_wsgi.

What mod_wsgi does is override the destruction sequence so that it will first go through each sub interpreter when and shutdown threading explicitly, then call atexit handlers. When that is done, only then will it destroy the sub interpreter and the main interpreter.

I have noted this previously in discussion associated with:

https://bugs.python.org/issue6531