There's scattered information in Slack, but I'm trying to make sense of it. What's the state of the art for dealing with packages that expect a dependency to be present in the environment at install or build time? Eg.

particles

expecting

numpy

in setup.py, or

torch-geometric

expecting

torch

at build time?

Is this a case where the package essentially has bad metadata and doesn't fully declare its dependencies? If so, you can use

python_requirements(overrides="...")

(or the similar overrides field in whatever target generator you use. https://www.pantsbuild.org/stable/reference/targets/python_requirements#overrides In short, the steps are: 1. Identify what package is needed (and version/extras/etc.) 2. Add this requirement to your project manually 3. add an override to tell pants that the

particles

requirement depends on

numpy

. Now, they should always be sandboxed together and things (should) work

I tried that, by setting

dependencies

in an overide but as I understand that applies to runtime deps, not buildtime. I also tried setting the

requirements

but I suspect that just replaces the package you want with the package you're setting.

hmm, yeah, I'm not sure about buildtime things. If that doesn't solve it at buildtime then I'm afraid I'm not too sure

np - Ive seen John Sirois pop up a few times to say "fork the package, build it, host a wheel" which ... I dunno, maybe that's an option, but it's a faff .