python / typing

Python static typing home. Hosts the documentation and a user help forum.
https://typing.readthedocs.io/
Other
1.59k stars 234 forks source link

Annotate sync/async code via Generic | generic TypeVar / Type aliases in Generic #1183

Open Bobronium opened 2 years ago

Bobronium commented 2 years ago

Consider this case:

import asyncio
from typing import Any, Coroutine, Generic, TypeVar

from httpx import AsyncClient, Client, Response

ClientT = TypeVar("ClientT", bound=Client | AsyncClient)

class API(Generic[ClientT]):
    client: ClientT

    def __init__(self, client: ClientT) -> None:
        self.client = client

    def get_items(self) -> Response | Coroutine[Any, Any, Response]:
        return self.client.request('get', 'https://example.com/api/v1/get_items')

response = API[Client](Client()).get_items()
response.json()  # Item "Coroutine[Any, Any, Response]" of "Union[Response, Coroutine[Any, Any, Response]]" has no attribute "json"  [union-attr]

response_coroutine = API[AsyncClient](AsyncClient()).get_items()
asyncio.run(response_coroutine)  # Argument 1 to "run" has incompatible type "Union[Response, Coroutine[Any, Any, Response]]"; expected "Awaitable[Response]"  [arg-type]mypy(error)

How can I express bound between ClientT and return type of API().get_items()?

How it could look like if Generic would support type aliases or TypeVar supported another type variables:

import asyncio
from typing import Annotated, Any, Awaitable, Coroutine, Generic, TypeVar
from typing_extensions import reveal_type

from httpx import AsyncClient, Client, Response

ClientT = TypeVar("ClientT", bound=Client | AsyncClient)
T = TypeVar("T")

# Using Annotated just as generic type that returns its argument as is
Sync = Annotated[T, ...]
MightBeAwaitable = TypeVar("MightBeAwaitable", bound=Awaitable | Sync)
# or MightBeAwaitable = Awaitable | Sync

class API(Generic[ClientT, MightBeAwaitable):
    client: ClientT

    def __init__(self, client: ClientT) -> None:
        self.client = client

    def get_items(self) -> MightBeAwaitable[Response]:  # TypeError: 'TypeVar' object is not subscriptable
        return self.client.request('get', 'https://example.com/api/v1/get_items')

response = API[Client, Sync](Client()).get_items()
reveal_type(response)  # Revealed type is "Response"

response_coroutine = API[AsyncClient, Awaitable](AsyncClient()).get_items()
reveal_type(response_coroutine)  # Revealed type is "typing.Awaitable[Any]"

Sorry if it's the wrong place/type for this issue.

Bobronium commented 2 years ago

Might be duplicate / use case of #548

relsunkaev commented 2 years ago

Perhaps something like this?

import asyncio
from typing import Any, Coroutine, Generic, TypeVar, overload

from httpx import AsyncClient, Client, Response

ClientT = TypeVar("ClientT", bound=Client | AsyncClient)

class API(Generic[ClientT]):

    def __init__(self, client: ClientT) -> None:
        self.client = client

    @overload
    def get_items(self: "API[Client]") -> Response: ...

    @overload
    def get_items(self: "API[AsyncClient]") -> Coroutine[Any, Any, Response]: ...

    def get_items(self) -> Response | Coroutine[Any, Any, Response]:
        return self.client.request('get', 'https://example.com/api/v1/get_items')
Bobronium commented 2 years ago

Interesting. It could work, I'll try it. Thank you!

Though, does it imply that every method needs to be annotated as like thus, or it can be done only for root methods and return annotations for ones that use them can be omitted?

relsunkaev commented 2 years ago

This would have to be done for every method that becomes sync/async based on the type of client passed in to __init__.

Bobronium commented 2 years ago

Then I'd say its too much of a duplication/overloads that going to obstruct actual code. Writing it in .pyi files will help with readability, but still will make process of writing/changing the code more complicated than it should be.

relsunkaev commented 2 years ago

There really isn't a way to correctly type this without overloads. The solution suggested at the end of the question wouldn't really work, even if HKTs did make it into Python, as it would allow for something like

response = await API[Client, Awaitable](Client()).get_items()

without the type checker complaining. This is equivalent to having a response_type parameter and passing in Awaitable or Sync. It doesn't actually enforce anything.

P.S.: I would also switch Awaitable for Coroutine since that is more "correct". If some decorator requires a Coroutine as return type, it will not accept get_items methods since Awaitable is not compatible with Coroutine, even though the code actually works. Generally, try to be as broad as possible on input types and as precise as possible on output types.