From #425 (comment)
address already in use failures
This one's tricky. Getting rid of the hard-coded port (e.g. NOT_54321_PORT = get_unused_port()) would for sure get rid of this, but it might be indicative of a deeper cleanup problem... perhaps the atexitasync stuff isn't firing properly, or some other async shenanigans? Perhaps it's not interacting correctly with pytest's freaky stuff?
From #425 (comment)