You can easily do that by simply running N instances and distributing requests among them. No need for all the complication that would come with removing the GIL.

And pay in memory for that luxury: when using lots of complex libraries (like in ML) the Python process can eat quite a lot of RAM.

Since gc.freeze() got implemented (https://docs.python.org/3/library/gc.html#gc.freeze) having forked workers has much lower cost these days. As long as you can load your data pre-fork that is.

Wow, I missed its introduction. Nice. On a cursory glance Celery doesn't make use of it yet, unfortunately.

Yes. Python is not without tradeoffs.