stop using the json API at pypi

[dimbleby]

This is intended mostly as a simplification: now that the pep658 backfill is complete we have two ways of accessing essentially the same information, and it is better to have just one. This also allows the relevant code to be common among all HttpRepository classes.

This should be a small win for any packages with no dependencies. For these packages: poetry reads the json API, correctly mis-trusts it when it says that the package has no dependencies and then ends up at the metadata anyway. Now we just go straight to the metadata.

Otherwise I suspect this is largely a wash: most of the time it swaps one network request for another, and instead of parsing a json response it parses a pkg-info.

To be honest I might not have started on this if I had known how tedious sorting out the tests was going to be...

The test script changes were larger than hoped mostly because it turns out the pypi-repository had a "fallback" flag. The meaning of this flag seems to be "behave correctly": but it is set to False in unit tests. The refactor eliminates that flag and as a result the unit tests now find themselves using more fixtures - because dependencies are correctly being chased.

I have removed two testcases rather than patch them up:

• test_fallback_pep_658_metadata() becomes somewhat redundant since lots of tests are now using metadata
• test_solver_skips_invalid_versions()
  • This test was added at 1ebd629, the code that it then tested survives I think only in PypiRepository.search() - which is not what the test covers at all
  • this one testcase was pulling in a large number of fixtures in the pandas / numpy / scipy / matplotlib family, I judge that it just is not worth it

[abn]

This is great work @dimbleby. A few notes;

1. Can we please add the metadata file fetching into https://github.com/python-poetry/poetry/blob/main/tests/repositories/fixtures/pypi.org/generate.py so that we can manage them going forward.
2. Lets remove the description from the files as we do not really use them in our tests.
3. The binaries added in dist should really be added by the generate script emptying out unnecessary content and updating hashes in relevant files.

[dimbleby]

I had a look at generate.py: but it did not work for me and when I tried to fix it it did not do the things that I wanted it to do. Fixture updates here were made without it.

I am not motivated to update it to do the work that I have already done. My view would be that if it was useful for you at #9132: super, but I am not interested in updating it. If you think that it is likely to be useful to you then I have no objection to you updating it...

(fwiw my method for downloading metadata files to fixtures was to add some temporary code in the mocking: ie I arranged that when a testcase wanted a missing fixture, it just added it).

removing descriptions etc messes with the hashes, IMO this sort of thing is just generating unnecessary work compared to using the real files.

[abn]

I am not motivated to update it to do the work that I have already done. My view would be that if it was useful for you at #9132: super, but I am not interested in updating it. If you think that it is likely to be useful to you then I have no objection to you updating it...

I can appreciate that it is added effort for this sort of change, but as maintainers it does remove certain pain points and allows us to manage the sprawl of randomly added files. While it might not be "useful" for you or "interesting" to you, it does make life easier for us.

(fwiw my method for downloading metadata files to fixture was to add some temporary code in the mocking: ie I arranged that when a testcase wanted a missing fixture, it just added it).

I think that is a great way of doing it. The original list of things I used in #9132 was similarly generated, but opted for a static list in generate.py as it wont lead to tests succeeding when it shouldn't etc. But yes, for discovery and filling items it makes a lot of sens to do it this way.

[dimbleby]

If updating generate.py is something that you want then I am not going to try to stop you. But "useful" or "interesting" are pretty much exactly my deciders for poetry contributions and my judgement is that this is neither of those things.

I would be willing to update this MR to remove generate.py, I acknowledge that the merge has left us in an inconsistent place where the file exists but does not do what it promises.

[abn]

As this change is in the right direction we want to go as a project, I am sure this will eventually get across the line.

The following, in my opinion, will need to be resolved prior to merging this.

1. The mocked test fixtures should be generated such that it can be regenerated along with the existing fixtures - ie. the generate.py script must be updated.
2. The fixtures in dist should only be for distributions that are fully preserved for the purposes of the test. Rest must be stubbed.
3. Optionally, any data that is not required for the purposes of the test must be removed.

If you do not want to make those changes yourself; feel free to leave this PR for another contributor to carry/address the code review comments.

[dimbleby]

I am sceptical about the value of "stubbing" files in dists. The files added there in this pull request are - combined - considerably smaller than the existing setuptools wheel: the extra complication that stubbing causes in making hashes consistent across metadata and scripts is to my mind not at all worthwhile for the small number of bytes saved.

Anyway I think I probably have been clear enough but so as to leave no doubt: I do not agree that the effort/reward ratio in updating generate.py is favourable and I do not intend to do it.

[abn]

I am sceptical about the value of "stubbing" files in dists. The files added there in this pull request are - combined - considerably smaller than the existing setuptools wheel: the extra complication that stubbing causes in making hashes consistent across metadata and scripts is to my mind not at all worthwhile for the small number of bytes saved.

I can appreciate your skepticism. The existing wheels are there as they have been required by some test cases - ideally I would like to remove them as well. The main intent for stubbing the files has nothing to do with saving bytes (it is a good side-effect), we want to avoid hosting and distributing usable binaries and especially ones that contain vulnerable code. The modified hashes also has a useful property, albeit not required at the moment, of making sure we are testing against the fixtures and not accidentally fetching files from the internet. The hash computation steps also allow us to add custom fixtures and generate hashes in links we care about.

Anyway I think I probably have been clear enough but so as to leave no doubt: I do not agree that the effort/reward ratio in updating generate.py is favourable and I do not intend to do it.

Understood. We simply see it differently.

Requested changes were implemented in #9175

[abn]

@dimbleby I will fix up the generate script, there are a few bad assumptions made that need fixing.

[abn]

@dimbleby made the required changes at #9186 feel free to pick them up from there.

One thing I noticed while getting that working was that this change forces Poetry to pull sdists in the cases a package does not distribute wheels. Which surprisingly seems to be a fair few projects. What this leads to is an awkward situation where the average user's UX is likely impacted negatively with this change.

I am no longer sure if removing it entirely is the best course of action until sdists are also populated. Although, I should also point out that fetching the sdist and inspecting is likely going to more yield more accurate results than the JSON API based on previous issues handled.

cc: @python-poetry/core

[dimbleby]

this change forces Poetry to pull sdists in the cases a package does not distribute wheels.

that is not changed by this pull request. Project dependencies on the json API are populated only from wheels (specifically, they are populated only if the first uploaded distribution for a version is a wheel). Therefore poetry already has to pull the sdist to determine dependencies for sdist-only releases and this is unavoidable.

what has changed is that previously the unit tests set fallback = False on the PyPiRepository. The meaning of this flag seems to have been "pretend that the json API is reliable when it says no-dependencies", but since that is a bogus assumption the flag was never set outside of unit tests.

[abn]

The fallback = False case is only one instance of this. For example, https://pypi.org/pypi/trackpy/0.6.2/json lists the requires_dist value, my understanding (granted I have not looked at the code path much recently) is that the values here are respected presently as it is not empty. This in-turn means we download the sdist that otherwise we would not have.

[dimbleby]

I am surprised to see useful values for an sdist-only release.

Perhaps I am out of date about how things work - it looks as though something changed recently to make this possible: https://pypi.org/pypi/trackpy/0.6.1/json does not provide requires_dist but https://pypi.org/pypi/trackpy/0.6.2/json does.

I'm not sure whether I agree that this is enough reason to stick with the json API, but I do agree that the balance is different than I thought it was an hour ago!

I haven't seen pypi provide metadata for any source distributions, only for wheels. Does that agree with your experience? Maybe worth a report on their issue tracker to ask whether this is expected or fixable.

[abn]

I'm not sure whether I agree that this is enough reason to stick with the json API, but I do agree that the balance is different than I thought it was an hour ago!

I think others on core proposed that we might consider keeping the JSON API as a fallback option if we detect an sdist only scenario. That will mean keeping the logic around longer.

I haven't seen pypi provide metadata for any source distributions, only for wheels. Dose that agree with your experience? Maybe worth a report on their issue tracker to ask whether this is expected or fixable.

My understanding was that the metadata would only be available for wheels. The pep is a bit ambiguous on that I believe. So a question might be worth posing.

[dimbleby]

My understanding was that the metadata would only be available for wheels

I think that's right, I think it is returning METADATA rather than PKG-INFO.

I think my view would be that the json API is sufficiently un-approved by the warehouse folk, and that sdist-only packages are relatively infrequent, and we should go ahead anyway and encourage anyone who is annoyed by the slowdown to go and bug maintainers to publish wheels. But I'm not 100% committed to that view

As I said at the start, this pull request has altogether been more trouble than I had hoped! I doubt I am likely to work up the enthusiasm any time soon to rearrange it so that it partially reinstates the json api: if y'all think that is a good direction then please don't wait on me to do it!

[dimbleby]

notes on recent commits:

• thanks for the work on generate.py.
• that the "test fixture consistency" test is passing only shows that it does not spot all the files that have been deleted
• I've continued to remove test_fallback_pep_658_metadata(), it is surely made redundant now that pretty much everything uses pep 658 metadata
• I've tweaked test_find_packages_does_not_select_prereleases_if_not_allowed() so that it uses a package where the fixtures actually list some prereleases

re sdists and the json API: I haven't understood what, if anything, has changed recently.

But I have been keeping half an eye out; and have not encountered any other examples where an sdist-only release has its requirements listed on that API (but several examples where they do not).

I still do not expect to do anything about the sometimes-use-the-json-api-after-all idea, and remain content that not using it is reasonable.

[dimbleby]

A couple of recent examples of the extra confusion that we cause by using the JSON API at pypi - and therefore introducing code paths for pypi that are quite different from the non-pypi case.

The failure to parse metadata 2.3 with older pkginfo is masked when reading dependencies from the json API:

• on the plus side, this issue would be more severe if it affected pypi users
• however it is much more confusing for users who encounter it only on mirrors or other non-pypi repositories. Much harder, for example, to provide a repro when the supposed repro works just fine against the major public repository.
• there probably ought already to have been a poetry patch release requiring pkginfo 1.10, I am pretty sure that there would have been if pypi users were impacted

The mis-publishing of docutils 0.21 as described in #9293, #9297

• we get different behaviour depending on whether docutils is found on pypi or, say, a mirror
• again, causing extra confusion

I expect there are and will be more, perhaps I will update this comment as I find them.

No knockout blow saying that "we must get off the JSON API" here, just adding to the accumulation of reasons that it would be helpful.

[Secrus]

@edgarrmondragon how is PEP 833 connected with this PR?