[matrss]

That's what I mean, in this case pip, uv, etc. are the wrong tool to use. You could e.g. use pixi and install all python and non-python dependencies through that, the conda-forge package of shapely will pull in geos as a dependency. Pixi also interoperates with uv as a library to be able to combine PyPI and conda-forge packages using one tool.

But conda-forge packages (just like PyPI packages, or anything that does install-time dependency resolution really) are untestable by design, so if you care for reliably tested packages you can take a look at nix or guix and install everything through that. The tradeoff with those is that they usually have less libraries available, and often only in one version (since every version has to be tested with every possible version of its dependencies, including transitive ones and the interpreter).

All of these tools have a concept similar to environments, so you can get the right version of GEOS for each of your projects.

[amelius]

Indeed, I'd want something where I have more control over how the binaries are built. I had some segfaults with conda in the past, and couldn't find where the problem was until I rebuilt everything from scratch manually and the problems went away.

Nix/guix sound interesting. But one of my systems is an nVidia Jetson system, where I'm tied to the system's libc version (because of CUDA libraries etc.) and so building things is a bit trickier.

[dagw]

with uv (and pip) you can pass the --no-binary flag and it will download the source code and build all you dependencies, rather than downloading prebuilt binaries.

It should also respect any CFLAGS and LDFLAGS you set, but I haven't actually tested that with uv.

[amelius]

I just tried --no-binary with the torchvision package (on a Jetson system). It failed. Then I downloaded the source and it compiled without problems.