Improve variant priorisation for mamba

[wolfv]

We're currently trying to fix mamba resolutions when variant packages are involved. It already works reasonably well when (directly) requiring a package that has a track feature applied. However, it stops working well when we have variants that do not directly expose any track_feature information.

For example, lets say we have 5 numpy builds:

• numpy = 1.20 = py36
• numpy = 1.20 = py37
• numpy = 1.20 = py38
• numpy = 1.20 = pypy36
• numpy = 1.20 = pypy37

The metadata for numpy numpy=1.20=py37 contains a dependency on python >=3.7,<3.8. However, the metadata for numpy=1.20=pypy37 contains exactly the same python dependency. The dependencies differ in python_abi and the pypy37 variant package has an additional pypy3.7 dependency.

However, the de-priorization by track_feature is applied on the python package. This makes it hard to pick the preferred solution right away, since we cannot figure this out without doing a full resolution.

In my opinion it would be preferable to either:

• inherit track_features into the numpy variants to make sure that pypy packages are down-weighted. Nothing prevents us from adding track_features into the numpy-variant package as well, and I would say that it would be "more correct" to do so. However, a downside is that this might increase complexity, especially for the conda solver which does additional computations with the track_features that we do not take into account in mamba.
• We could specifically pin the python package via build string to the pypy variant. If we'd export a stricter dependency (e.g. python >=3.7,<3.8 *pypy) for numpy's pypy build, we would be able to "inherit" the de-priorization by inspecting the first level of dependencies. This should be a fairly quick process. If we'd get consensus for this idea, we could have a rule that a variant package must indicate a de-priorization in the first level of dependency via build-string pinning (or on the package itself if the variants depend on differently named packages).

Does conda-forge think that's reasonable? I think for the case of pypy we could do some pretty straight-forward repodata patches to add *pypy to all packages that also require pypy3X.

[isuruf]

I really don't like option 1 because of added complexity involved in the recipes. Option 2 is doable. I wonder if this should be fixed in libsolv or not. How does this affect which version is used? For eg: conda prioritizes python 3.9 over 3.8.

[wolfv]

Thanks for commenting! :)

Yes, we're working on improvements in libsolv:

• the first improvement is to order variants by their dependencies (or rather dependency boundaries). The plan is to find the lower-, and upper bound and then prefer the one with the higher lower bound. In certain cases we might also have to compare the upper bound, we'll see. We could also down-weight packages with 'more' dependencies but not sure if that is going to work across the board (would work in the pypy case as then numpy has the additional pypy37 dependency).
• the second improvement would be to inspect variant dependencies that have a build string to check if they exclusively select packages with a track-feature, and if yes, down-weight them.

[adriendelsalle]

Thanks @wolfv

What is the rationale making a package with a higher lower bound preferable vs higher higher bound?

For the second possible improvement, this is not limited to solvables that have a build string isn't it? It just that in python case that build string will only select providers that have a track feature.

[isuruf]

we could have a rule that a variant package must indicate a de-priorization in the first level of dependency via build-string pinning

I think python is a special case where there were 2 run exports python and python_abi. For everything else, this rule should apply. For python, de-prioritization was indicated in python and the variant was selected using python_abi. I've fixed that in https://github.com/conda-forge/python_abi-feedstock/pull/11