typeshed
typeshed copied to clipboard
python
[stubsabot] Bump protobuf to 4.21.*
Release: https://pypi.org/project/protobuf/4.21.3/ Homepage: https://developers.google.com/protocol-buffers/
If stubtest fails for this PR:
- Leave this PR open (as a reminder, and to prevent stubsabot from opening another PR)
- Fix stubtest failures in another PR, then close this PR
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
Hmm, it appears actions do not trigger unless you close and open the PR manually :-/
As mentioned in https://github.com/python/typeshed/issues/8403#issuecomment-1197251076 , a major version change is a good opportunity to drop the __init__.py , see https://github.com/python/typeshed/issues/7519
As mentioned in #8403 (comment) , a major version change is a good opportunity to drop the
__init__.py, see #7519
It also sounds like pyright might be considering patching the protobuf stubs they bundle: https://github.com/microsoft/pyright/issues/3751
Pyright patching it is good for users that don't install types-protobuf (maybe is majority).
If someone has types-protobuf in there dev-requirements they'll override bundled stubs and be confused. My understanding of non stdlib stubs is users are recommended to specify them as a dependency? You could treat protobuf as special and recommend rely on bundled ones, but I'd guess it'd confuse most users given it's a fairly niche bug.
Pyright patching it is good for users that don't install types-protobuf (maybe is majority).
If someone has types-protobuf in there dev-requirements they'll override bundled stubs and be confused. My understanding of non stdlib stubs is users are recommended to specify them as a dependency? You could treat protobuf as special and recommend rely on bundled ones, but I'd guess it'd confuse most users given it's a fairly niche bug.
Ah that's a v good point. We should probably make the change here regardless of what pyright does with respect to patching the bundled version, in that case, then.
For the record I didn't force-push anything. I think it's because I created the SSH key that stubsabot now uses.
I think the plan is to make use the opportunity of the major version bump of protobuf removes the __init__.pyi. This caused a lot of problems when we tried it earlier / users mentioned it would have been a little less surprising with a major version bump. And then removing the __init__.pyi is least disruptively done once mypy 0.990 is out (because it changes the default of --namespace-packages).
I tried deleting google/__init__.py, but mypy doesn't seem to like it (at least not in our CI), despite --namespace-packages now being the default in mypy 0.990. Anybody have any idea what's going on there?
I reproed. This appears to have to do with --explicit-package-bases
https://mypy.readthedocs.io/en/stable/running_mypy.html#mapping-file-paths-to-modules
I was able to get things to pass by adding --explicit-package-bases to the tests, and adding MYPYPATH=stubs/protobuf. This is probably not an ideal solution.
It is probably a bug inside mypy regarding how --custom-typeshed-dir works with relation to namespace packages.
Nipunn is correct that we need to use --explicit-package-bases and add all stubs to MYPYPATH (or copy files to a different kind of directory structure and check them). I don't think there's an interaction with --custom-typeshed-dir, that should just be used for stdlib stubs these days (and maybe mypy_extensions).
Nipunn is correct that we need to use
--explicit-package-basesand add all stubs to MYPYPATH (or copy files to a different kind of directory structure and check them).
Will that also mean that the stubs package will be of no use to users who don't use the --explicit-package-bases option, though? That seems like it might be quite disruptive (but maybe that's okay, if we're doing a major version bump here?). Is there any possibility of --explicit-package-bases becoming the default in the future? Does it have any downsides?
[...]
--custom-typeshed-dir[...] should just be used for stdlib stubs these days (and maybe mypy_extensions).
Hmm, I disagree somewhat. From time to time we have PRs to our stdlib stubs that cause mypy to fail on our third-party stubs (e.g. in #8908, the change to stdlib/datetime.pyi caused mypy to start failing on our third-party stubs for python-dateutil, meaning that those stubs had to be altered in the same PR). I'd much rather we got notified of those failures in the same PR. If we didn't use --custom-typeshed-dir for our third-party stubs, mypy would have suddenly started to fail on our stubs for python-dateutil when we next bumped mypy, due to a typeshed PR relating to the stdlib datetime stubs two months prior that nobody remembered anymore.
This shouldn't have any impact on users who install stubs. The issue here is that in our CI we pass some filenames to mypy and mypy needs to convert filenames to module names somehow.
Yup, I agree with that. I was trying to respond to Nipunn's suggestion that there was a bug in --custom-typeshed-dir. mypy doesn't use --custom-typeshed-dir to find third party stubs (other than mypy_extensions maybe).