Brains should be dynamically registered

[nickdrozd]

Current problem

There is an Astroid function for detecting Numpy members (attribute_looks_like_numpy_member). I'm sure it is very useful for code that uses Numpy, but for code that doesn't use Numpy it is a big waste of time. For example, running Pylint against Pylint itself, this function is called 1,441,536 times, and it is never true because Pylint doesn't use Numpy.

This makes a measurable time difference.

Desired solution

For a given Astroid brain, some effort should be made to detect whether that brain is actually needed. This might entail doing a pass over imports, or something like that. For a codebase that does not use Numpy, the Numpy brain should not be registered and none of its checks should be run, and similar for other brains.

Additional context

No response

[DanielNoord]

I agree with not calling code that is never useful, however this won't add too much to performance.
I looked into this before: since the numpy brain is the first that is registered we do a lot of inference work in it. By not loading it we just move that work to a different brain without noticeably improving performance.

[Pierre-Sassoulas]

I think this would be a good thing, but it's not easy to do. I think we'd have to go from a "battery included" pylint that "just works" for numpy, to a lot of plugin with optional install pylint[numpy,spelling] or at least option if the configuration, all solutions that requires user actions during an upgrade. (There's not only numpy, we have brains for dozen of libs). I don't know how much numpy 2.0 will break the numerous numpy brains for 1.0 we have, so an user action might be necessary anyway (or do we autodetect the numpy version?).