Change versioning scheme to be more robust #57
Conversation
Currently sqlalchemy breaks typeshed CI basically whenever it releases a new micro version. This would allow us to pin sqlalchemy to a tighter version in METADATA.toml, while ensuring users who do `pip install types-sqlalchemy` will still get the latest stub. Please see the comments in `compute_incremented_version` about what we attempt to guarantee (and the corresponding asserts in that function). There's also some associated refactoring so that we can unit test logic easily and so that we can lean a little more heavily on `packaging` to do things for us.
| # We'll try to bump the fourth part of the release. So e.g. if our version_spec is | ||
| # 1.1, we'll release 1.1.0.0, 1.1.0.1, 1.1.0.2, ... | ||
| # But if our version_spec is 5.6.7.8, we'll release 5.6.7.8.0, 5.6.7.8.1, ... |
There was a problem hiding this comment.
Why change this behavior? For example, types-tzlocal currently has version 4.2.2, while tzlocal has version 4.2.
There was a problem hiding this comment.
I think it helps disambiguate "upstream version" from "types version". I've seen someone be confused by this and pin unnecessarily strictly.
E.g. if tzlocal released 4.2.1 that would still be compatible with 4.2 specified in METADATA.toml. But the correct corresponding version of types-tzlocal is still 4.2.2.
The issue comes about because the target version in METADATA.toml is hidden information to the average user. Very few packages use the fourth spot, so defaulting to the fourth spot in practice makes it clearer that the version isn't 1:1 with upstream versions since they are unlikely to overlap.
Is this a change you're open to?
There was a problem hiding this comment.
I'm fine, but I'm concerned this could break types-tzlocal (and potentially others). I would like to see this changed in a separate PR to keep possible breakage per PR minimal.
There was a problem hiding this comment.
I'm curious if your concern can be turned into suggestions for tests. I just added a test_version_increment that runs determine_incremented_version on all distributions we publish. This checks that all invariants we assert are met for all currently distributed packages. Let me know if there's anything else that would increase confidence in these changes.
That aside, yes, I can comment this change out for now and raise exception in the change above
|
Thanks for the review!! |
Co-authored-by: Sebastian Rittau <[email protected]>
Co-authored-by: Sebastian Rittau <[email protected]>
|
Bump, I think I've made all requested changes + added more tests |
|
I'm sorry, I have no time to review, but please merge if you and @JelleZijlstra are satisfied. And thanks for working on this! |
|
Merging since I'm not sure @hauntsaninja has merge access here. |
This change was removed from typeshed-internal#57, so as to make that PR more a refactoring. The changes here accomplish two things: First, allows us to change the specificity of how we pin versions in typeshed while ensuring that users who install without pinning versions still get the latest stub. Second, incrementing the fourth position helps disambiguate "upstream version" from "types version", reducing user confusion. See my comment here: typeshed-internal#57 (comment)
This change was removed from #57, so The changes here accomplish two things: First, allows us to change the specificity of how we pin versions in typeshed while ensuring that users who install without pinning versions still get the latest stub. Second, incrementing the fourth position helps disambiguate "upstream version" from "types version", reducing user confusion. See my comment here: #57 (comment)
Currently sqlalchemy breaks typeshed CI basically whenever it releases a
new micro version. This would allow us to pin sqlalchemy to a tighter
version in METADATA.toml, while ensuring users who do
pip install types-sqlalchemywill still get the latest stub.Please see the comments in
compute_incremented_versionabout what weattempt to guarantee (and the corresponding asserts in that function).
There's also some associated refactoring so that we can unit test logic
easily and so that we can lean a little more heavily on
packagingto dothings for us.