I dabble with python occasionally and I'm always fighting with tools and tool combinations that don't really combine well. The last time I settled on using conda to get some isolation of python versions and then pipenv for getting some sane package management with a lock file. Not pretty but it kind of worked. Except I had a hard time convincing vs code and pycharm of the correct environment with that combination (couldn't resolve libraries I installed). I got it working eventually but it wasn't a great experience.
It sounds like uv should replace the combination. Of course there is the risk of this being another case of the python community ritually moving the problem every few years without properly solving it. But it sounds like uv is mostly doing the right thing; which is making global package installation the exception rather than the default. Most stuff you install should be for the project only unless you tell it otherwise.
Will give this a try next time I need to do some python stuff.
Do. I was sceptical at first - exactly because of the points you make: I mostly do ML, so for getting PyTorch and Cuda etc. to play nice conda was basically my go-to.
We use poetry at work, but getting it to play nice with PyTorch is always a bit of an art.
I tried to get into Pixi, but have been a little annoyed as it seems to have inherited conda's issues when mixing conda and PyPi.
It sounds like uv should replace the combination. Of course there is the risk of this being another case of the python community ritually moving the problem every few years without properly solving it. But it sounds like uv is mostly doing the right thing; which is making global package installation the exception rather than the default. Most stuff you install should be for the project only unless you tell it otherwise.
Will give this a try next time I need to do some python stuff.