Distribute typing stubs as part of Traits

[mdickinson]

When we have two separate packages being released from a single repository, release management and versioning are unnecessarily painful. If we want to keep traits-stubs as a separate package, we may want to consider breaking it out into its own repository.

The alternative would be to fully absorb traits-stubs into traits.

[rahulporuri]

Personally, i'm for fully absorbing. I felt overhead when i was looking to add/modify stubs because they are two separate packages at the moment. It will also mean easier maintenance of the stubs - otherwise, every API change in traits would need a separate change in the stub.

[mdickinson]

I have both philosophical and practical objections to fuller integration of typing in Traits. Let's leave aside the philosophical objections for now. On the practical side, typing.py and traits.api define a lot of the same names with related but different meanings: e.g., Dict, List, Any, Union. I think this is going to be a recipe for confusion if we want to use these in type annotations.

That's partially fixed in 3.9, where the "list of integers" type can be spelled as list[int] rather than List[int], but it's going to be a while before we can drop 3.8 as being too old.

Integrating just to the point where traits-stubs and traits are part of the same distribution, share the same version, changelog, etc., seems possible.

[mdickinson]

Integrating just to the point where traits-stubs and traits are part of the same distribution, share the same version, changelog, etc., seems possible.

I think a prerequisite for such integration would be improving the traits-stubs quality and completeness: if installing traits from PyPI (or EDM) also automatically installs traits-stubs, it's harder for mypy-using users to opt out of using the traits stubs. We've got a way to go before the traits stubs are of usable quality, but right now they're opt-in rather than opt-out, so that's okay.

[mdickinson]

From offline discussion, keeping traits-stubs in this repository and having those stubs automatically distributed with Traits seems like the way to go. We have a couple of options here:

• keep traits-stubs as a separate package in the same distribution
• move the .pyi files into the traits package

NumPy recently moved stubs into the core package, it may be worth looking at how they're managing this.

[mdickinson]

Updated the title to reflect what we actually want to do.