Adjust path_to_url et al. to produce the same results on Python 3.14+ #13423
Conversation
| return urllib.request.pathname2url(urllib.request.url2pathname(part)) | ||
| ret = urllib.request.pathname2url(urllib.request.url2pathname(part)) | ||
| if sys.version_info >= (3, 14): | ||
| ret = ret.removeprefix("//") |
There was a problem hiding this comment.
I am not proud of this :(
There was a problem hiding this comment.
It wasn't obvious to me how else to do this either 🙁
There was a problem hiding this comment.
Is there something in cpython we can reference here so people reading this later knows what’s going on?
There was a problem hiding this comment.
I can try digging a reference to the change.
There was a problem hiding this comment.
Ideally, if we could find something that explained why cpython changed what they did, and why the new result is "correct" and why pip needs something different from the "correct" answer, that would be best (as it would give us a much better basis for informed decisions if this code ever needs to change again).
Unfortunately, I get the impression that there's no real "correct" answer here, and the cpython change was "because it's more consistent with (something or other that pip maybe doesn't even care about)". If so, then documenting what precisely pip is using to base its idea of what "the url for a pathname" is, would be better than nothing.
Worst case scenario would be that there's simply no standard for how to convert a pathname to a URL in general, and it's all just a mess of guesswork and hacks 🙁
There was a problem hiding this comment.
According to this comment it is this is a result of CPython's attempt to follow RFC 3986, Section 5.2.2: python/cpython#125974 (comment)
I beleive python/cpython#125989 and python/cpython#123273 are the related PRs.
Previous discussions on the "standards" around path to url: #13062 (comment)
There was a problem hiding this comment.
I didn't get to this today. If you want, feel free to adjust this PR with some comments.
There was a problem hiding this comment.
My main point here is that as a pip maintainer, the removeprefix('//') line looks like complete magic - there's no clear reason why it's needed, why it's 3.14+ only, or why urllib.request isn't doing the right thing for us in the first place. And I'm not particularly comfortable with a comment that essentially just says that we're compensating for a change in CPython, because that begs the question why are we having to effectively revert the CPython change, rather than raising this as a bug on the stdlib function. I'd like to see something that explains why it isn't a CPython bug, and why in that case we need to alter the behaviour.
Having said all of this, I don't think it's fair to expect @hroncok to do all that research and analysis, so if we have to go with a comment that says "CPython changed behaviour in a way that broke our code, so we're patching the result back to what we expect" then I'll live with it.
@barneygale as the author of #13062 (comment) can you shed any light on this?
There was a problem hiding this comment.
I wonder if having a custom wrapper around just urllib.request.pathname2url wouldn't be clearer. Call it e.g. path2url_fragment and describe what it should return and that it cannot start with //.
|
With no other feedback, are you planning to remove draft status from this and add a small news item? I tested this locally and it appears to be good, but because it's slightly changing the behavior of brittle path functions I think there should be a small bug fix note. |
|
Can do. |
|
Anyway. I added a news entry and a link to the upstream discussion to a comment. Marked as non-draft. That's about as much as I have the capacity to do. |
|
With my RM hat on, barring any major issues (I still need to review this), I plan on including this in the upcoming 25.2 release regardless of whether CPython replies or not. I'd like to include it in the release so downstream can use/package pip for the Python 3.14 betas. |
See python/cpython#125974 and #13138 (comment)
This makes the tests pass on Fedora with Python 3.14.0b2 (except test_get_index_content_directory_append_index, which segfaults python/cpython#135448)