CI: Fix slow mamba solver issue by limiting boto3 versions

[datapythonista]

Closes #61531

Probably better to rerun the CI 3 or 4 times to be sure this is the problem and the solution. But based on local tests, seems like boto3 has a huge number of versions (they release almost every day), and that's the problem with the mamba solver. Limiting the number of versions provided to the solver should help. 1.27 is from 2 years ago, consistent with other packages. Why only fails for 3.13? No idea

[datapythonista]

This doesn't seem to work. Removing boto3 from the environment seems to fix the problem locally, I thought limiting the options would also fix it. But seems like even pinning to a specific version doesn't work. I guess the problem may be with a boto3 dependency, but not too sure.

[datapythonista]

The dependencies for boto3 are here: https://github.com/conda-forge/boto3-feedstock/blob/main/recipe/meta.yaml#L23

Not too sure if the problem is really with boto3, maybe it's more complex but removing boto3 makes it work. But doesn't seem like mamba really checks their issues, and I don't think we want to get into the mamba solver internals ourselves, so I guess we need to just keep trying something that makes the solver for 3.13 take a reasonable time.

[datapythonista]

Based on my tests, this is what makes the solver happy. Relaxing the conditions causes the very slow mamba install times. This breaks the check for minimum versions, so it can't be merged as is (and I'll add a comment when we decide what to do).

@mroeschke not sure if you have an opinion on what to do to avoid the mamba problems based on this. Not too sure what's going on.

[mroeschke]

I'm curious if conda's libmamba solver has the same issue. Potentially based on conda/conda-libmamba-solver#668 it may. (It's on my radar to maybe switch to https://github.com/conda-incubator/setup-miniconda given that mamba activity is kinda slow as you mentioned)

It's a shame we have to pin, but I'm OK with it to speed up solve times

[datapythonista]

I had the impression that conda, mamba, micromamba and pixi all use libmamba, so not sure if changing tool would be helpful. But I'm not so sure about the exact details.

[WillAyd]

I don't know what conda is doing with its solver, but I've seen this issue with boto3 and the standard pip dependency resolver, which downloads a copy of every possible boto version and inspects its metadata to find out what constraints it might have to solve for. The sheer number of boto versions to download makes that a really slow process

[datapythonista]

I'll leave out the s3fs>=2025.5.1 for now, and just pin boto3 in all the dependency files. I see mamba being very slow in other versions of pandas now. Since boto3 is not a pandas dependency, but something we use for testing, pinning shouldn't be a problem. Just pining boto3 didn't make mamba resolve immediately in my tests, but with some luck it makes resolve faster, and we can avoid the cancelled jobs.