uv fails to resolve apache-airflow[all]==2.8.4 where pip succeeds
#3078
Comments
|
One issue could be pip has a preference on what to backtrack on based on certain heuristics of the users requirements, see:
Whereas my understanding is uv does not apply a preference like this. |
konstin
added
the
compatibility
|
We added the heuristics in #3087 but this still fails for me, for a different reason related to the relationship between extras and the base package. I have some ideas on how to fix it. |
Yeah, I was only speculating that pip's heuristics are part of it. With regards to extras, pip has an optimization that "constrains" extras together with their base requirements: pypa/pip#12095, don't know if that approach will be helpful to you, as no idea how uv handles extras, but for some use cases it was a big performance boost for pip. Also, my understanding is pubgrub-rs strongly implements CDCL to resolve, whereas I would describe pip's approach as "weakly" implementing CDCL, so the path taken through the resolution graph will always be a little different. Though in this case I would think uv's approach would be better, and I do hope to eventually improve pip here and make it's approach to CDCL stronger. |
|
Unfortunately I still can't get this to avoid going down the |
I won't have time today, but I'll try and make a simpler set of requirements that reproduces this issue. |
|
Simplifying uv fails with the requirement When I look at the metadata of Spending a lot of time playing around with requirements, I found the issue was a combination of Edit: Constrained failing requirements even further, and deleted the below comment as I wasn't able to reproduce the results this morning and I don't want people to take it as true if I can't reproduce it. Further I want to investigate why pip skips over |
|
Okay, I think I've found why pip succeeds here and uv fails, and it's not going to be intuitive. So, currently, pip has a preference on "backtrack causes": https://github.com/pypa/pip/blob/24.3.1/src/pip/_internal/resolution/resolvelib/provider.py#L192 But what are backtrack causes? You might assume that they are the causes that have disjoint requirements, but it turns out resolvelib is sufficiently generic that it doesn't have a concept of disjoint. It just has a concept of that a cause was known to be involved in a failure. So for example, let's say you have A depends on D > 1, B depends on D > 2, and C depends on D < 2, then resolvelib considers A, B, and C to all be causes, even though A's requirements aren't disjoint from either B or C. The effect of this is that pip sort of "wanders" around the possible requirements when there are more than 2 possible causes, resulting in it sometimes being able to find an "unintuitive" solution at the cost of it sometimes being very slow. If I make resolvelib "disjoint aware": pypa/pip@main...notatallshaw:pip:speedy-resolve, then pip ends up getting exactly the same error. As such, I think I'm going to close this issue, as I think uv is doing the right thing here. |
Streps to reproduce
uv pip install --dry-run apache-airflow[all]==2.8.4Error
Expected
Same as pip, can install fine:
The text was updated successfully, but these errors were encountered: