consider thinning supported matrix

[minrk]

petsc4py and slepc4py now have 260 build variants because of:

• 2 mpi providers
• 2 scalar variants
• 6 Python versions
• 2 mac architectures
• 3 linux architectures
• 3 cuda variants on linux

This is causing rerenders to take at least 30 minutes, and time out on the bot, so I don't think any automated PRs are going to be made to these repos until something changes.

I think we should consider culling some of these to keep things manageable. Options include:

• drop Python 3.8 altogether (44 builds)
• only build cuda12 on Python >=3.11 (36 builds)
• only build cuda at all on Python >=3.10 (48 builds) or 3.11 (72 builds)
• only keep 'complex' builds on Python >=3.10 (44 builds)

It would really be ideal if petsc4py didn't need to be aware of the cuda build variants at all, but I think @stefanozampini confirmed that it doesn't work that way (I can't find the reference anymore, though).

[stefanozampini]

Unfortunately it's not possible at the moment to use a petsc at runtime which is different than the build one

[dalcinl]

@stefanozampini Could you please remember me why we have 3 CUDA variants?

[stefanozampini]

The cuda variants are 11.8, 12.4 and nocuda

[minrk]

Looks like the rerender times are fixable upstream, and thinning the build matrix doesn't actually help rerender time. More info here: https://github.com/regro/cf-scripts/issues/2561 and a couple linked conda-build PRs.

So we should only bother thinning here if build time is problematic.

[minrk]

closing this since rerender time is a lot more manageable now