Support for attrs dataclasses

[bilelomrani1]

attrs is a popular third-party package for defining dataclasses with additional convenient features (converters, more fine-grained initialization, etc.). Currently simple-parsing only supports native dataclasses, is there any interest to support attrs dataclasses as well?

[lebrice]

Hello there @bilelomrani1, thanks for posting!

Hmm that's interesting. Does attrs have an API that is interchangeable with the built-in dataclasses?

For parser.add_arguments(<attrs.dataclass>, dest="foo") to work, what we'd need are equivalents to:

• dataclasses.fields(dataclass_type_or_instance) -> Iterable[dataclasses.Field]
• dataclasses.is_dataclass(dataclass) -> bool
• dataclasses.Field attributes:
  • name: str
  • type: str | type
  • default: Any | dataclasses.MISSING
  • default_factory: Callable[[], Any] | dataclasses.MISSING
  • metadata: dict

If there's a way of get this information from the attrs.dataclass, then yeah, this would be feasible, nice contribution! :)

[lebrice]

Hey @bilelomrani1, do you know if attrs has this kind of interop with dataclasses ?

[bilelomrani1]

Hi @lebrice I'm sorry for the long delay. I no longer need this feature as I found an existing alternative for my use case. But I can still provide some answers if you are still interested in this feature, I once wrote a dataclasses/attrs parser some time ago, the code is available here.
Looking at it, here is what I used:

• attr.fields(attr_type): list of attrs fields, containing the field name, type, factory/default value.
• attr.has(attr_type) --> bool: equivalent of dataclasses.is_dataclass(dataclass) -> bool

[chicham]

Hi @lebrice,

I'm interested in this integration too!

For your questions:

Hmm that's interesting. Does attrs have an API that is interchangeable with the built-in dataclasses?

Yes, absolutely!

dataclasses.fields(dataclass_type_or_instance) -> Iterable[dataclasses.Field]

attrs.fields(cls) -> tuple[attrs.Attribute, ...]

dataclasses.is_dataclass(dataclass) -> bool

attrs.has(cls) -> bool

dataclasses.Field attributes:
name: str
type: str | type
default: Any | dataclasses.MISSING
default_factory: Callable[[], Any] | dataclasses.MISSING
metadata: dict

https://www.attrs.org/en/stable/api.html#attrs.Attribute

For the record, the dataclass module is heavily inspired by the attrs library, and attrs is usually more featured with a very close API.

I'd be happy to help explore this further! What are your thoughts?

[lebrice]

Sure, we can add support for attrs dataclasses. A first step could probably be to replace all the uses of the functions from dataclasses with generic functions with the same name, that dispatch either to the attrs function or dataclasses function depending on the class /field type.

If you're willing to help with that, I'd be happy to give some pointers or feedback!