It's not the case anymore since 1b356b8c because it was broken. The issue is now about re-implementing threading in the cli :P

I implemented such features using greenlets, it's not too fancy and super-reliable. How does it sound ?

https://greenlet.readthedocs.io/en/latest/

what do you think ?

It does rely on live-patching subproc.call and run though. Probably not what we want.

Pool and map is the other option. Provide a demo of what exif externals would expect then ?

8 forks is a safe default, should mat2 be -j compliant ?

Greenlet is written in C, which causes several issues in my opinion:

• It adds a binary dependency
• Its seems not-so-portable
• If it's not packaged, people would have to compile it themselves
• It's a bit overkill, in my opinion, to bring in a native module and some boilerplate, to gain a tiny bit of speed

What is wrong with using stuff that are already in Python's stdlib?

I didn't consider the bin-deps. Multiproc Pool is the way to go then.

Done in 88b95923