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
D
Replying to @mitsuhiko
Every single data scientist uses compiled packages and that will never change because Python is not an appropriate language to *implement* algorithms in. The design of wheels, statically linking in all dependencies is fundamentally inferior to conda packages because of binary and memory size.
15,000 views
M
Replying to @dhirschfeld
That represents the status quo
15,000 views
D
Replying to @mitsuhiko
Let's all just drop wheels and move to conda packages. IMO everyone would be better off.
15,000 views
M
Replying to @dhirschfeld
I greatly disagree. Also that is not very likely for a range of reasons.
15,000 views
D
Replying to @mitsuhiko
I think pixi is wise to enable safe mixing of wheels and conda packages rather than insist on conda packages. For Pure Python packages, there is no harm in wheels. @mitsuhiko.at I cannot imagine a path for data science that does not need compiled code and something like conda. What do you imagine?
15,000 views
M
Replying to @danielballan
There is an implication here that packages distributed on pypi cannot contained compiled code and I'm puzzled why that is. At the end of the day what you need is a resolver that installs wheels, that's it. The rest is just a question of creating wheels and hosting them.
15,000 views
D
Replying to @mitsuhiko
The key is “statically linking in all dependencies”.
15,000 views
M
Replying to @danielballan
Wheels do not need to statically link all dependencies
15,000 views
Replies
D
Replying to @mitsuhiko
And then they're not reproducible because they depend on the vagaries of the system installed packages. You can copy a coda environment to a distroless container. AFAIK that's impossible with wheels because they depend on the system they were compiled in.
15,000 views
E
Replying to @mitsuhiko
Conda/pixi do solve real problems for some users, for instance because there is currently no way for a wheel to express its non-python dependencies. There is a lot of good information on complexities of non-python dependencies here: pypackaging-native.github.io/key-issues/n...
15,000 views