homepage

threading._shutdown() uses threading.enumerate() which iterations on threading._active.
threading.Thread registers itself into threading._active using its _bootstrap_inner() method. It unregisters itself when _bootstrap_inner() completes, whereas its is_alive() method still returns true: since the underlying native thread still runs and the Python thread state still exists.
_thread._set_sentinel() creates a lock and registers a tstate->on_delete callback to release this lock. It's called by threading.Thread._set_tstate_lock() to set threading.Thread._tstate_lock. This lock is used by threading.Thread.join() to wait until the thread completes.
_thread.start_new_thread() calls the C function t_bootstrap() which ends with:
 tstate->interp->num_threads--;
 PyThreadState_Clear(tstate);
 PyThreadState_DeleteCurrent();
 PyThread_exit_thread();
_PyThreadState_DeleteCurrent() calls tstate->on_delete() which releases threading.Thread._tstate_lock lock.
In test_threads_join_2() test, PyThreadState_Clear() blocks on clearing thread variables: the Sleeper destructor of the Sleeper instance sleeps.
The race condition is that:
* threading._shutdown() rely on threading._alive
* Py_EndInterpreter() rely on the interpreter linked list of Python thread states: interp->tstate_head.