are the generic types worth it

[d-v-b]

I am not convinced that using generics for GroupSpec and ArraySpec is worth it.

It's nice to be able to write

MyArray = ArraySpec[AttrsModel]

but I suspect many people find this confusing and would get more out of examples like

class MyArray(ArraySpec):
    attributes: AttrsModel

In principle the generic types could bring value for automatically generating return types of methods like GroupSpec.to_flat, but we can't really resolve the return value of to_flat because it collects everything from subgroups, and the type annotations of those subgroups are not easy to access.

So maybe it would be worth exploring a branch without the generic types. What would feel worse? I'm not sure! But I know a few things that would certainly feel better (like not needing to specialize GroupSpec all the time).

[d-v-b]

there might also be a middle ground, where we keep some generic typing for the members of a group, but we treat attributes like a plain-old attribute

[d-v-b]

to expand on this: i am 99% sure we get nothing by making the attributes of an arrayspec / groupspec a type parameter. Until we can use type defaults for type parameters, the requirement a type parameter for attributes adds needless boilerplate to using these classes. So I would support removing that type parameter.

[tlambert03]

will parrot here what I said on zulip. I agree with you @d-v-b that "many people find this confusing and would get more out of examples like". I too think it's pretty nifty that you can express this sort of thing with generics: "this is a thing that will always have an attributes field... but I don't know what that field will be just yet". But since it's so common to just define fields with types in pydantic (as opposed to parametrizing generic models), it probably is a bit more surprising this way.