Guido> So in order to do correct diagnostics here we would have to add an "owning
thread" pointer to every event loop;

I didn't understand why this approach was not taken when the check was introduced. I was surprised that get_event_loop() was used for the check instead.

Here is a patch checking the thread using threading.get_ident(). The check still works when set_event_loop(None) is used in unit tests.

I created the issue #23074 for the change: get_event_loop() must always raise an exception, even when assertions are disabled by -O.

--

asyncio requires thread support, it is already the case without my patch. For example, the event loop policy uses threading.Local.

In the aioeventlet project, call_soon() writes a byte in the self pipe (to wake up the selector) if the selector is waiting for events; call_soon() is "green thread safe". It is less efficient than calling forcing users to call explicitly call_soon_threadsafe(), but aioeventlet is written to be compatible with asyncio and eventlet, while these projects are very different.

In the aiogevent and aioeventlet projects, I also store the current greenlet, but for a different purpose: ensure that yield_future() is not called in the greenlet running the event loop, which would blocking.