Post

40 replies
D
That's fundamentally impossible because uv is limited to installing Python packages. If all you want is a faster pip uv is a fine choice. If you want a package manager that can install robust, reproducible data science environments then you need conda packages and the pixi package manager.
8,837,200 views
10
M
Replying to @dhirschfeld
My hope is that for data scientists that use Python the need to ever use conda goes to zero :)
15,000 views
J
Replying to @mitsuhiko
Conda gives you two things: 1. Language agnostic environment management 2. The ability to build more complex packages with compiled dependencies Assuming we only care about Python, uv has given us really nice environment management, and wheels have helped with build dependencies.
15,000 views
J
Replying to @jacobtomlinson
There's still a lot of work to be done to catch up with conda. Optional constraints, virtual packages and shared compiled libraries to name a few. But it's not crazy to imagine these things arriving over the next few years.
15,000 views
M
Replying to @jacobtomlinson
Most of those are not technical things but social ones. There is nothing that would stop you from publishing shared native libraries to PyPI today. That it's not done is in part I think because the need is fulfilled by conda.
15,000 views
D
Replying to @mitsuhiko
I guess in that world pip install openmp becomes a thing? and you’d get libgomp.so or libomp.dyib depending on Linux or macOS and then some C extension that links to it can drop using auditwheel/delocate and declare openmp a dependency?
15,000 views
J
Replying to @ddavis
Do we need pypi-forge?
15,000 views
D
Replying to @jacobtomlinson
Python developers will love having one more place to distribute packages! 😵‍💫
15,000 views
M
Replying to @ddavis
PyPI will work fine
15,000 views
J
Replying to @mitsuhiko
Conda Forge is a community of dedicated packagers more than a technical solution (although they have built impressive CI infrastructure for building packages). PyPI would work fine as the place to put the files. But we will need a PyPI-Forge community to build them.
15,000 views
M
Replying to @jacobtomlinson
Yes and no. I think there is an element to what forge does today but i don’t think that third parties should build Python packages. However to distribute non Python shared libraries via PyPI would be beneficial and would need people.
15,000 views
J
Replying to @mitsuhiko
Yeah there are a lot of non-Python things on Conda Forge. But there are also a lot of things there that are depended on by Python packages. Those would need to be packaged on PyPI. But maybe Nix is the solution to reproducible environments if PyPI isn’t going to be that.
15,000 views
M
Replying to @jacobtomlinson
There is a tendency to over-complicate things. The goal should be that a user can always use uv to install packages directly from PyPI and for that to install binaries created by the authors of those packages. If those want to band together to depend on shared libraries then this is good.
15,000 views
J
Replying to @mitsuhiko
That’s a great goal! But this thread started with “let’s unite under uv and remove the need for conda/pixi”. So much of this discussion has been around the value that the conda ecosystem provides today. It’s dismissive of the needs of the science community to call those things overcomplicated.
15,000 views
Replies
M
Replying to @jacobtomlinson
I’m not sure why you’re getting those vibes. I said I want to see a future where conda is unnecessary and I stand by that. That does not mean conda does not have uses or that it does not provide solutions for users.
15,000 views