Pip doesn’t really work. That is, it works sometimes, but instead of getting a global view of the different versions of dependencies, it just starts downloading and installing things and it can get itself stuck in a corner.

Some developers use pip all the time. I have some little projects that I do in PyCharm and I add packages with the IDE which uses pip and it is fine.

I worked at a place where our projects were pretty complicated and pip could not solve them consistently.

My RSS reader YOShInOn uses conda for package management because CUDA “just works” with conda plus you get numpy and the like compiled against the MKL libraries for a nice boost. Conda is slower than it should be, I think it was a serious mistake that they compress packages with BZ2 (I had a system that even installed a local copy of all the CUDA libs with conda packages I made but decompression just takes forever.). Still, the solver is logically sound.

The worst problem of it is that the packages you want might not be in conda, you can get away with pipping in a package or two but long term that’s the road to ruin. If you do want to use packages that aren’t in conda your beat way forward is to build conda packages for them which is usually not hard —- to really get control of the build of any egg packages you need to make wheels out of them anyway which is basically the same.

It is a little ugly but YOShInon has a sidecar application which doesn’t have any machine learning in it but instead links the mastodon API client and other things that aren’t in conda so YOShInOn just runs that executable when it wants to post to Mastodon.