> In bpo-33608 I moved the pending calls to per-interpreter state.  We saw failures (sometimes sporadic) on a few buildbots (e.g. FreeBSD) during runtime finalization.  However, nearly all of the buildbots were fine, so it may be a question of architecture or slow hardware.  See bpo-33608 for details on the failures.

That was a crash in multiprocessing tests. PyEval_RestoreThread(tstate) was called with a freed PyThreadState: tstate->interp = 0xdbdbdbdbdbdbdbdb for example.

I recently fixed PyEval_RestoreThread() in bpo-39877 to exit properly daemon theads without dereferencing tstate which points to freed memory.

To reproduce the bug, it helped me to add a sleep at Python exit. Fixed value (ex: 1 second) or better: random sleep (between 1 ms and 1 sec).