[QUESTION] How to use with async?

[tiangolo]

[QUESTION] How to use with async?

[tiangolo]

Hello all! Thanks for all the discussion and workarounds.

I want to add support for async in Typer, without making it required or default, using AnyIO as an optional dependency. In FastAPI/Starlette, AnyIO is required, as everything is just async underneath. But in Typer I don't want to force people to use AnyIO when they don't really need it. That makes the whole thing a bit more complex as I need to add all the logic to conditionally use or not AnyIO and do async or not.

This is one of the top priorities for Typer, along with a code reference (API reference). I'm finishing some things in FastAPI, I have some things in SQLModel too, and then I'll continue with Typer stuff.

[gpkc]

I like @ryanpeach solution. I've made an alternative that doesn't require patching the Typer class directly. You can just inherit from it normally:

from asyncio import run
from functools import wraps

import typer


class AsyncTyper(typer.Typer):
    def async_command(self, *args, **kwargs):
        def decorator(async_func):
            @wraps(async_func)
            def sync_func(*_args, **_kwargs):
                return run(async_func(*_args, **_kwargs))

            self.command(*args, **kwargs)(sync_func)
            return async_func

        return decorator


app = AsyncTyper()

@app.async_command()
async def my_async_command():
    ...

@app.command()
async def my_normal_command():
    ...

if __name__ == "__main__":
    app()

[ryanpeach]

Actually the decorator @cauebs wrote doesn't make sense to me (maybe I just misunderstand click and anyio). The point is to support running the asynchronous function in two modes:

1. Called directly from the cli
2. Called as an asynchronous function from another function as a library function

If you just decorate the function with a function that makes it synchronous, you've ruined it.

But also we need argument information preserved.

So I propose the following:

# file: root/__init__.py
from functools import wraps
from asyncio import sleep, run
import typer

# This is a standard decorator that takes arguments
# the same way app.command does but with 
# app as the first parameter
def async_command(app, *args, **kwargs):
    def decorator(async_func):

        # Now we make a function that turns the async
        # function into a synchronous function.
        # By wrapping async_func we preserve the
        # meta characteristics typer needs to create
        # a good interface, such as the description and 
        # argument type hints
        @wraps(async_func)
        def sync_func(*_args, **_kwargs):
            return run(async_func(*_args, **_kwargs))

        # Now use app.command as normal to register the
        # synchronous function
        app.command(*args, **kwargs)(sync_func)

        # We return the async function unmodifed, 
        # so its library functionality is preserved
        return async_func

    return decorator

# as a method injection, app will be replaced as self
# making the syntax exactly the same as it used to be.
# put this all in __init__.py and it will be injected into 
# the library project wide
typer.Typer.async_command = async_command

# file: root/some/code.py
import typer
from asyncio import sleep
app=typer.Typer()

# The command we want to be accessible by both 
# the async library and the CLI
@app.async_command()
async def foo(bar: str = typer.Argument(..., help="foo bar")):
    """Foo bar"""
    return await sleep(5)

if __name__=="__main__":
    app()

This is written in such a way it could be literally written as a PR and put as a method into typer.main.Typer.

Thoughts?

[jessekrubin]

@neimad1985 Easier, but less fancy than the decorator solution is to just nest your async func:

from asyncio import run as aiorun

import typer


def main(name: str = typer.Argument("Wade Wilson")):
    async def _main():
        typer.echo(f"Hello {name}")


    aiorun(_main())


if __name__ == "__main__":
    typer.run(main)

[ryanpeach]

Noted but I don't think I've been impolite in this thread... And I believe I've added a relevant feature (the ability to wrap async functions as cli commands). Correct me if I'm wrong.

[jessekrubin]

@neimad1985

from asyncio import run as aiorun

import typer


async def _main(name: str):
    typer.echo(f"Hello {name}")

def main(name: str = typer.Argument("Wade Wilson")):
    aiorun(_main(name=name))


if __name__ == "__main__":
    typer.run(main)

[aogier]

i'm not able to be enthusiast about your attitude in this PR, this is where my emoji stem from. Given the irrelevant value your little boilerplate adds upstream this will neither add nor remove value to this library in my (irrelevant) opinion.

[gerardpc]

Any news on the async support? Any way we can help?

[gilcu2]

AsyncTyper for anyio (Still with typing problems)

import inspect
from functools import partial, wraps

import anyio
import asyncer
import typer
from typer import Typer


class AsyncTyper(Typer):
    @staticmethod
    def maybe_run_async(decorator, f):
        if inspect.iscoroutinefunction(f):

            @wraps(f)
            def runner(*args, **kwargs):
                return asyncer.runnify(f)(*args, **kwargs)

            decorator(runner)
        else:
            decorator(f)
        return f

    def callback(self, *args, **kwargs):
        decorator = super().callback(*args, **kwargs)
        return partial(self.maybe_run_async, decorator)

    def command(self, *args, **kwargs):
        decorator = super().command(*args, **kwargs)
        return partial(self.maybe_run_async, decorator)


app = AsyncTyper()


@app.command()
async def async_hello(name: str, last_name: str = "") -> None:
    await anyio.sleep(1)
    typer.echo(f"Hello World {name} {last_name}")


@app.command()
def hello() -> None:
    print("Hello World")


if __name__ == "__main__":
    app()

[csheppard]

Just found pallets/click#85 (comment) which may help me

[elpapi42]

We all need this

[cauebs]

@neimad1985 A decorator might help you:

from functools import wraps
import anyio

def run_async(func):
    @wraps(func)
    def wrapper(*args, **kwargs):
        async def coro_wrapper():
            return await func(*args, **kwargs)

        return anyio.run(coro_wrapper)

    return wrapper


@run_async
async def main(name: str = typer.Argument("Wade Wilson")):
    typer.echo(f"Hello {name}")

You can even have async completions:

import click

def async_completion(func):
    func = run_async(func)

    @wraps(func)
    def wrapper(*args, **kwargs):
        try:
            return func(*args, **kwargs)
        except (click.exceptions.Abort, click.exceptions.Exit):
            return []

    return wrapper


async def list_users() -> List[str]:
    ...


@run_async
async def main(
    name: str = typer.Argument("Wade Wilson", autocompletion=async_completion(list_users))
):
    typer.echo(f"Hello {name}")

[macintacos]

The workaround here are very clever! In my codebases I'm trying to use type annotations wherever possible, so I annotated the above to the point where I think everything works fine. If there are any type annotation aficionados here, feel free to rip it apart:

import asyncio
from collections.abc import Callable, Coroutine
from functools import wraps
from typing import Any, ParamSpec, TypeVar

import typer

P = ParamSpec("P")
R = TypeVar("R")


class AsyncTyper(typer.Typer):
    """Asyncronous Typer that derives from Typer.

    Use this when you have an asynchronous command you want to build, otherwise, just use Typer.
    """

    def async_command(  # type: ignore # Because we're being generic in this decorator, 'Any' is fine for the args.
        self,
        *args: Any,
        **kwargs: Any,
    ) -> Callable[
        [Callable[P, Coroutine[Any, Any, R]]],
        Callable[P, Coroutine[Any, Any, R]],
    ]:
        """An async decorator for Typer commands that are asynchronous."""

        def decorator(  # type: ignore # Because we're being generic in this decorator, 'Any' is fine for the args.
            async_func: Callable[P, Coroutine[Any, Any, R]],
        ) -> Callable[P, Coroutine[Any, Any, R]]:
            @wraps(async_func)
            def sync_func(*_args: P.args, **_kwargs: P.kwargs) -> R:
                return asyncio.run(async_func(*_args, **_kwargs))

            # Now use app.command as normal to register the synchronous function
            self.command(*args, **kwargs)(sync_func)

            # We return the async function unmodified, so its library functionality is preserved.
            return async_func

        return decorator

[thedrow]

It'd be nice if we could use Typer with the async-click fork.

[mreschke]

Seeing as how FastAPI is an async framework, having an async CLI seems logical. The main reason being sharing code from the CLI and Web entry points. You can of course use the asgiref.sync.async_to_sync converter helpers to call existing async methods from the CLI but there are complications here and it makes your cli code clunky. I replaced typer with smurfix/trio-click (which is asyncclick on Pypi) and it works great, but of course this is just async click, not the cool typer implementation. Forking typer and replaceing all import click with import asyncclick as click works like a charm but it means maintenance of the fork yourself. If @smurfix could keep his async fork maintained and up to date with click upstream, and if typer was based on asyncclick then we would really have something great here.

[mreschke]

Thanks @smurfix. Once you have asyncclick updated, if @tiangolo doesn't have a nice async typer by then, perhaps ill make a good typer-async fork of typer and use asyncclick and add to pypi for us all to use.

[hozn]

Also interested in a solution to this.

[toppk]

I think the solutions above by @macintacos @rafalkrupinski @gpkc are close, but they miss the magic of typer, which is no instructions needed. using inspect.iscoroutinefunction() (in python since 3.5) we can just do the code automatically so users can just do @app.command() as normal.

This implementation also allows to switch event loop (example show for uvloop) using the asyncio.Runner available in python >= 3.11.

I think something like this would be good to have in typer as it fits the simple design of typer.

#!/usr/bin/python

import typer
import asyncio
import inspect
import uvloop
import sys
from functools import wraps


class UTyper(typer.Typer):
    def __init__(self, *args, loop_factory=None, **kwargs):
        super().__init__(*args, **kwargs)
        self.loop_factory = loop_factory

    def command(self, *args, **kwargs):
        decorator = super().command(*args, **kwargs)
        def add_runner(f):
            
            @wraps(f)
            def runner(*args, **kwargs):
                if sys.version_info >= (3, 11) and self.loop_factory:
                    with asyncio.Runner(loop_factory=self.loop_factory) as runner:
                        runner.run(f(*args,**kwargs))
                else:
                    asyncio.run(f(*args,**kwargs))

            if inspect.iscoroutinefunction(f):
                return decorator(runner)
            return decorator(f)
        return add_runner

app = UTyper(loop_factory=uvloop.new_event_loop)


@app.command()
def hello(name: str):
    print(f"Hello {name}")


@app.command()
async def goodbye(name: str, formal: bool = False):
    if formal:
        print(f"Goodbye Ms. {name}. Have a good day.")
    else:
        print(f"Bye {name}!")
    await asyncio.sleep(1)

if __name__ == "__main__":
    app()

prints:

$ ./test.py goodbye asdf --formal
Goodbye Ms. asdf. Have a good day.
$ ./test.py hello asdf 
Hello asdf

and the 'goodbye' test pauses for a second, as expected.

[ryanpeach]

@cauebs I suppose jumping to a PR is a bit of a jump, I wasn't really aware the project was big enough to have a lot of maintainers or QA. We are planning on using the code I just wrote on a rather big company cli, so the feature is required for us, and is complete for our purposes. I just demo'd how to add basic anyio support, and I'll help out as best I can. I'm sure the PR can provide good scaffolding for a future solution.

[Butch78]

You could possibly implement Tiangolo's new little reop Asyncer, it has Aynio support for making functions async. It is very similar to @ryanpeach's implementation of anyio :)

https://github.com/tiangolo/asyncer

[wizard-28]

What is the final solution for this? I'm a bit confused.

[gghez]

is this project deprecated? There is no answer to the merge request for 1 year.

[toppk]

@borissmidt good find. yes, the entire loop factory/runner thing is not necessary. i've posted a concise version below.

#!/usr/bin/python

import typer
import asyncio
import inspect
import uvloop
from functools import wraps


class UTyper(typer.Typer):
    def command(self, *args, **kwargs):
        decorator = super().command(*args, **kwargs)
        def add_runner(f):
            
            @wraps(f)
            def runner(*args, **kwargs):
                asyncio.run(f(*args,**kwargs))

            if inspect.iscoroutinefunction(f):
                return decorator(runner)
            return decorator(f)
        return add_runner

app = UTyper()
app2 = UTyper()

@app2.command()
async def hello(name: str):
    print(f"Hello to subtest {name}")
    print(f"{asyncio.get_running_loop()=}")

app.add_typer(app2, name="sub")

@app.command()
def hello(name: str):
    print(f"Hello {name}")


@app.command()
async def goodbye(name: str, formal: bool = False):
    if formal:
        print(f"Goodbye Ms. {name}. Have a good day.")
    else:
        print(f"Bye {name}!")
    await asyncio.sleep(1)
    print(f"{asyncio.get_running_loop()=}")

if __name__ == "__main__":
    asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
    app()

[jonaslb]

The UTyper workaround above works well, but when used as a decorator it converts your coroutine functions into sync functions. This breaks calling a command from another command. I also think it is simply more appropriate to leave the function functionally unchanged, as typer already does for sync functions. The below adjustment should fix this, and will work even when calling the outer command:

class UTyper(typer.Typer):
    # https://github.com/tiangolo/typer/issues/88
    @override
    def command(self, *args, **kwargs):
        decorator = super().command(*args, **kwargs)

        def add_runner(f):
            if inspect.iscoroutinefunction(f):

                @wraps(f)
                def runner(*args, **kwargs):
                    return asyncio.run(f(*args,**kwargs))

                decorator(runner)
            else:
                decorator(f)
            return f

        return add_runner

app = UTyper()

@app.command()
async def inner():
    await asyncio.sleep(1)
    print("inner")

@app.command()
async def outer():
    await inner()


if __name__ == "__main__":
    app()

[uda]

I am working on a cli part of an async app with async DB connection, so needed the async cli as well, thanks for all contributors to the AsyncTyper / UTyper workaround, here is my working version:

import asyncio
import inspect
from functools import wraps, partial

from typer import Typer


class AsyncTyper(Typer):
    @staticmethod
    def maybe_run_async(decorator, f):
        if inspect.iscoroutinefunction(f):

            @wraps(f)
            def runner(*args, **kwargs):
                return asyncio.run(f(*args, **kwargs))

            decorator(runner)
        else:
            decorator(f)
        return f

    def callback(self, *args, **kwargs):
        decorator = super().callback(*args, **kwargs)
        return partial(self.maybe_run_async, decorator)

    def command(self, *args, **kwargs):
        decorator = super().command(*args, **kwargs)
        return partial(self.maybe_run_async, decorator)

I needed the callback to be async as well to handle database connection init instead of repeating it

[pythonwood]

pr #340 test ok. it is better.

from https://github.com/skeletorXVI/typer/tree/feature/anyio-support

not merge yet

good tutorial

import asyncio

import typer

app = typer.Typer()


@app.command()
async def wait(seconds: int):
    await asyncio.sleep(seconds)
    typer.echo(f"Waited for {seconds} seconds")


if __name__ == "__main__":
    app()

[rafalkrupinski]

class AsyncTyper(typer.Typer):
    def async_command(self, _func: Callable = None, *args, **kwargs):
        def decorator(async_func):
            @wraps(async_func)
            def sync_func(*_args, **_kwargs):
                return run(async_func(*_args, **_kwargs))

            self.command(*args, **kwargs)(sync_func)
            return async_func

        if _func:
            return decorator(_func)
        return decorator

Then it also works without parenthesis

@app.async_command
async def my_async_command():
    pass

edit: I just realised that original typer doesn't work without them :D

[rafalkrupinski]

Calling asyncio.run in the first line of the command does the job.

Of course you actually need to pass another function to asyncio.run, so for every command you need two functions, one of which is boilerplate.
So maybe you write a decorator for it... but it's such a common case, maybe someone wrote it already...
Wait, why it isn't a part of the library?

[gpkc]

I can't see any appeal for having it at the moment other than it being extra code to be maintained to be honest. What's the problem with just calling aiorun?

[JosXa]

I can't see any appeal for having it at the moment other than it being extra code to be maintained to be honest. What's the problem with just calling aiorun?

Either change the project motto to "Another random CLI lib that works a little bit like FastAPI" or introduce async and keep the "little brother of FastAPI" headline.

FastAPI without async? Huh?

Jokes aside, is this really such a large maintenance effort?

[smurfix]

I will update asyncclick to the latest click release as soon as anyio 2.0 is ready.

[amfarrell]

@mreschke I've updated @jessekrubin's PR to remove the conflicts with master, in case you find it useful.

[cauebs]

Okay, we're all trying to help here. Let's not take anything personally.

@ryanpeach Your solution is in essence very similar to mine, but yours is one step ahead. One thing you missed and that I will insist on is that we should tackle not only commands but also things such as autocompletion functions (and others I might be missing).

And another matter we should discuss before jumping to a PR (and here I kind of understand the discomfort you might have caused to @aogier) is supporting different async runtimes other than asyncio. I couldn't use the feature as it stands in your code, because I use trio instead of asyncio.

My proposal: add an "extra" on the package called anyio, alt-async or something else, that toggles an optional dependency on anyio=^3. Then, in this decorators impl, we check if anyio is available, and if it's not we fallback into asyncio (your current impl). Otherwise it's just a matter of time until someone opens another issue here requesting support for trio or something else.

On a final note, I usually wait for a maintainer to weigh in, so as not to waste any time on an unwelcome solution. I salute your initiative, but give people time to evaluate your proposal! 😄 Cheers.

[gpkc]

I don't think it makes sense for typer to have an async_command built in.
Calling asyncio.run in the first line of the command does the job.

Can you please point to any examples?

Just call asyncio.run inside the command if you want to do it like that.

from asyncio import run as aiorun

@app.command()
def my_command(email: str) -> None:
    async def do_my_command():
        from <some_module> import awaitable
        await awaitable()

    aiorun(do_my_command())

[borissmidt]

1 Looking at the python async docs the loop factory might be unnecessary if you set the eventloop policy. The disadvantage is that you have to set it before any app is ran. 2 An alternate is to make typer uvloop aware and set it when the package is installed. 3 The disadvantage of doing it on an typer.add is that it is a bit magic that the parents setting is ported to the child setting bur only for the loop. It would also be tricky to get it right with multiple levels of apps. Also when you then run the subapp separately you still don't get the uv eventloop. 4 maybe an classvar to set the eventloop can be used. I think option 2 needs buy in from the project owner. Option 1 seems the most correct solution so set the policy and i might need to update the pr to use asyncio.get_event_loop. Then the factory stays for app specific overrides. ```python asyncio.set_event_loop_policy(uvloop.EventLoopPolicy()) loop = asyncio.get_event_loop() ```

…

On Sat, 8 Jul 2023, 07:11 toppk, ***@***.***> wrote: Here's my attempt at some asyncio background: there is only one call to run during an event loop lifecycle. you can create tasks while an event loop is running, but if there isn't an event loop you need to start it up, and the run interface is the only game in town. The whole get_running_loop() interfaces is when you *know* there is a running loop and you want to do something with it. That's not applicable for us (well it is in the fact that I use it to demonstrate different event loops in use). Of course you can spin down the event loop and start up another one, but most of the time it's just one event loop per run. What typer does is route to a method, and we're putting in a asyncio.run (well specifically the 3.11 version of that interface) into the decorator. Since typer will never call two methods, then we're only starting one event loop. Getting the loop factory option shared between Typer instances linked via Typer.add_typer is probably a more obscure use case, because it's not clear how many people replace the event loop, and certainly how many people chain together Typer instances. But imagine if you have Typer instances started in a dozen modules then you have to go to each one to set the loop_factory, that then the one place where it all joins together. what I'm saying that it would be nice if you don't have to set the loop factory for every typer instance, just the root. This example below (that needs my UTyper class above, as well as uvloop package installed) shows that the loop_factory in use has to be set for each typer instance, and I think that's sub optimal, and unexpected behavior. Since this the add_typer call is after all the decorations done, such a solution will require some minor surgery that occurs at the time of the add_typer call. Hopefully not too bad. But if this work is too much, I think we can get it into a second PR after this one goes it (which is what I hope you're looking to do soon). It might be better to defer this, as it will certainly complicate the implementation. Demonstrating that the loop factor needs to be set to each Typer instance instead of just the root. $ python test-utyper.py goodbye foo Bye foo!asyncio.get_running_loop()=<uvloop.Loop running=True closed=False debug=False> $ python test-utyper.py sub hello foo Hello to subtest fooasyncio.get_running_loop()=<_UnixSelectorEventLoop running=True closed=False debug=False> app = UTyper(loop_factory=uvloop.new_event_loop)app2 = UTyper() @app2.command()async def hello(name: str): print(f"Hello to subtest {name}") print(f"{asyncio.get_running_loop()=}") app.add_typer(app2, name="sub") @app.command()def hello(name: str): print(f"Hello {name}") @app.command()async def goodbye(name: str, formal: bool = False): if formal: print(f"Goodbye Ms. {name}. Have a good day.") else: print(f"Bye {name}!") await asyncio.sleep(1) print(f"{asyncio.get_running_loop()=}") if __name__ == "__main__": app() — Reply to this email directly, view it on GitHub <#88 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABNXZFTB4GASOBGDK2W7W4TXPDTXFANCNFSM4MGSXYGA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[rafalkrupinski]

Nobody would be complaining about missing async support if was just for compatibility (similarity actually) with FastAPI. There's a increasingly common case of calling async functions, IMO the main example would be remote requests with httpx.
I must admit I'm really surprised with the resistance over the 15 lines of code this change represents.

[todofixthis]

I ran into a related problem when running async unit tests with pytest-asyncio because by the time the test calls runner.invoke(...) there is already a running event loop. I'm not aware of a straightforward way to work around this issue, but I managed to find a slightly convoluted one, by effectively doing the same async-to-sync trick with the test 🤓

Would love your feedback on whether there is a better way to go about this.

# cli/commands/generate.py
...
@app.command("profiles")
@embed_event_loop
async def generate_profiles():
    """
    Generates profile data using Random User Generator API (https://randomuser.me/) and
    adds them to the database.
    """
...

# test/integration/cli/test_generate.py
...
def test_generate_profiles_happy_path(runner: CliRunner):
    result = runner.invoke(["generate", "profiles"])
    assert result.exception is None
    ...
    @embed_event_loop
    async def verify():
        # Verify that the profiles were saved correctly to the database.
        async with db_service.session() as session:
            ...
    verify()
...

where @embed_event_loop is the "async-to-sync" decorator described in previous comments on this issue.

[mreschke]

Thanks guys. Ill need some time to pull it all in and prototype this instead of asyncclick. If this all works out what is the probability of merging this request and making it a part of this official typer repo. Optional async would be perfect. I really hate to fork permanently.

[killswitch-GUI]

I agree with @mreschke, we tightly couple all of our code and actually use Type CLI to call our uvicorn/guinicorn using various "management" commands. Ran into this once we wanted to use some of the async calls we have.

[jessekrubin]

Hi :)
Anything new on this ?

@neimad1985 I don't think async is PR-ed in yet, but I use async with typer all the time by just running the async processes from within my sync functions once the parsing is done. It works for most basic things.

[ryanpeach]

Tested the code I posted above and it works. You could probably just add it as a method to typer.Typer. I'll make a pr.

[ryanpeach]

Unsure why @aogier downvoted. It runs and its integrating well into my repo.

[NikosAlexandris]

The workaround here are very clever! In my codebases I'm trying to use type annotations wherever possible, so I annotated the above to the point where I think everything works fine. If there are any type annotation aficionados here, feel free to rip it apart:

import asyncio
from collections.abc import Callable, Coroutine
from functools import wraps
from typing import Any, ParamSpec, TypeVar

import typer

P = ParamSpec("P")
R = TypeVar("R")


class AsyncTyper(typer.Typer):
    """Asyncronous Typer that derives from Typer.

    Use this when you have an asynchronous command you want to build, otherwise, just use Typer.
    """

    def async_command(  # type: ignore # Because we're being generic in this decorator, 'Any' is fine for the args.
        self,
        *args: Any,
        **kwargs: Any,
    ) -> Callable[
        [Callable[P, Coroutine[Any, Any, R]]],
        Callable[P, Coroutine[Any, Any, R]],
    ]:
        """An async decorator for Typer commands that are asynchronous."""

        def decorator(  # type: ignore # Because we're being generic in this decorator, 'Any' is fine for the args.
            async_func: Callable[P, Coroutine[Any, Any, R]],
        ) -> Callable[P, Coroutine[Any, Any, R]]:
            @wraps(async_func)
            def sync_func(*_args: P.args, **_kwargs: P.kwargs) -> R:
                return asyncio.run(async_func(*_args, **_kwargs))

            # Now use app.command as normal to register the synchronous function
            self.command(*args, **kwargs)(sync_func)

            # We return the async function unmodified, so its library functionality is preserved.
            return async_func

        return decorator

Would you have any examples to share?

[borissmidt]

I agree, in the pr i made i only wrap the callback registered to typer. Not the function itself. I'll see if there was any progress on getting it merged.

I'm just stuck with the pr because the mypy configurations is to strict and I can't make it happy for all python versions 😞

[borissmidt]

Its a missing feature and i have a PR open but there is a funny collision between mypy and the python versions. #633 I also didn't get any response on it.

…

On Sun, 3 Sept 2023, 10:31 Joscha Götzer, ***@***.***> wrote: Is there any good reason that async commands are *not* part of the library? Or is it just a missing feature? — Reply to this email directly, view it on GitHub <#88 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABNXZFXA6D4L7DSO5IVYLZ3XYQ54RANCNFSM4MGSXYGA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[warrenbhw]

any news on this?

[jessekrubin]

Thanks @smurfix. Once you have asyncclick updated, if @tiangolo doesn't have a nice async typer by then, perhaps ill make a good typer-async fork of typer and use asyncclick and add to pypi for us all to use.

@mreschke I made a pull request to this repo that gets you most of the way to async. #128

[neimad1985]

Hi :)
Anything new on this ?

[neimad1985]

Thanks for the quick answer @jessekrubin
Would it be possible that you share a simple code example on how you do this please ?

[neimad1985]

@jessekrubin

Ok thanks, that's exactly what I was thinking.
The problem is the duplication of functions main and _main and their arguments.
If you have multiple subcommands, which I have, for your program you have more and more duplication.
Anyway thanks for answering me.

[neimad1985]

@cauebs
Thanks, that is actually a nice idea, I will try it !

[neimad1985]

@jessekrubin
Very nice trick. I should have thought about this. Thank you.

[septatrix]

Just found pallets/click#85 (comment) which may help me

As that issue was closed a few years ago and is now locked I decided to open a new one containing a bit more information and addressing some comments in the previous issue. You can look at it yourself, leave some feedback (preferably typer agnostic) and upvote it to show interest in this feature: pallets/click#2033

[ryanpeach]

I think just adding those decorators to the library and having @app.command() auto detect if the function it's decorating is async or not and just pick the appropriate decoration. Not hard at all to implement. Thanks everyone for the suggestions

[oramirite]

Typer mentions being able to stack the FastAPI and Typer decorators on top of each other because they're "meant to work together". Without async support this actually isn't possible though. I hope this PR will be accepted to provide a path towards a solution, even if imperfect, since the stated design goals of Typer are to work with FastAPI which is async in nature.

[AsgrimS]

To the people scrolling down to find a solution. In my opinion, what @neimad1985 did is really the simplest and works fine.

@neimad1985 Easier, but less fancy than the decorator solution is to just nest your async func:

from asyncio import run as aiorun

import typer


def main(name: str = typer.Argument("Wade Wilson")):
    async def _main():
        typer.echo(f"Hello {name}")


    aiorun(_main())


if __name__ == "__main__":
    typer.run(main)

If you don't want to nest functions you can also just:

import asyncio

import typer

def main(arg_1: str):
   asyncio.run(my_async_func(arg_1))

async def my_async_func(arg_1: str):
   await asyncio.sleep(5)

if __name__ == "__main__":
   typer.run(main)

[paketb0te]

@gpkc This is f*cking genius!

@tiangolo What do you think?

[Kludex]

I don't think it makes sense for typer to have an async_command built in.

Calling asyncio.run in the first line of the command does the job.

[NikosAlexandris]

I don't think it makes sense for typer to have an async_command built in.

Calling asyncio.run in the first line of the command does the job.

Can you please point to any examples?

[takeda]

This looks good, maybe this should transform into a PR?

[toppk]

Looks like someone is working on a PR, maybe.

One issue I'll leave here for @borissmidt is that if one were to use

subapp=Typer(loop_factory=...)

app=Typer(loop_factory=...)
app.add_typer(subapp,"mysub"))

then it would be nice that the "subapp" app would use the loop_factory from the root Typer() object. I hadn't written that up yet, and currently it doesn't do that in my child class or your native implementation AFAICT

[borissmidt]

I'm a bit junior in python asyncio, so if you have a sample that would be great. I expected that it is automatically used from the task context? I.e. you can call get_current_loop or something like that. I coded it in a way that if the async function is called directly that the loop factory isn't used.

…

On Fri, 7 Jul 2023, 21:03 toppk, ***@***.***> wrote: Looks like someone is working on a PR, maybe. One issue I'll leave here for @borissmidt <https://github.com/borissmidt> is that if you use subapp=Typer(loop_factory=...) app=Typer(loop_factory=...)app.add_typer(subapp,"mysub")) then it would be nice that the "subapp" app users the loop_factory from the root Typer() object. I hadn't written that up yet, and currently it doesn't do that in my child class or your native implementation AFAICT — Reply to this email directly, view it on GitHub <#88 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABNXZFR4C5OG2X2VYY33EQDXPBMRNANCNFSM4MGSXYGA> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[toppk]

Here's my attempt at some asyncio background: there is only one call to run during an event loop lifecycle. you can create tasks while an event loop is running, but if there isn't an event loop you need to start it up, and the run interface is the only game in town. The whole get_running_loop() interfaces is when you know there is a running loop and you want to do something with it. That's not applicable for us (well it is in the fact that I use it to demonstrate different event loops in use).

Of course you can spin down the event loop and start up another one, but most of the time it's just one event loop per run.

What typer does is route to a method, and we're putting in a asyncio.run (well specifically the 3.11 version of that interface) into the decorator. Since typer will never call two methods, then we're only starting one event loop. Getting the loop factory option shared between Typer instances linked via Typer.add_typer is probably a more obscure use case, because it's not clear how many people replace the event loop, and certainly how many people chain together Typer instances. But imagine if you have Typer instances started in a dozen modules then you have to go to each one to set the loop_factory, that then the one place where it all joins together.

what I'm saying that it would be nice if you don't have to set the loop factory for every typer instance, just the root. This example below (that needs my UTyper class above, as well as uvloop package installed) shows that the loop_factory in use has to be set for each typer instance, and I think that's sub optimal, and unexpected behavior.

Since this the add_typer call is after all the decorations done, such a solution will require some minor surgery that occurs at the time of the add_typer call. Hopefully not too bad. But if this work is too much, I think we can get it into a second PR after this one goes it (which is what I hope you're looking to do soon). It might be better to defer this, as it will certainly complicate the implementation.

Demonstrating that the loop factor needs to be set to each Typer instance instead of just the root.

$ python test-utyper.py goodbye foo
Bye foo!
asyncio.get_running_loop()=<uvloop.Loop running=True closed=False debug=False>

$ python test-utyper.py sub hello foo
Hello to subtest foo
asyncio.get_running_loop()=<_UnixSelectorEventLoop running=True closed=False debug=False>

app = UTyper(loop_factory=uvloop.new_event_loop)
app2 = UTyper()

@app2.command()
async def hello(name: str):
    print(f"Hello to subtest {name}")
    print(f"{asyncio.get_running_loop()=}")

app.add_typer(app2, name="sub")

@app.command()
def hello(name: str):
    print(f"Hello {name}")


@app.command()
async def goodbye(name: str, formal: bool = False):
    if formal:
        print(f"Goodbye Ms. {name}. Have a good day.")
    else:
        print(f"Bye {name}!")
    await asyncio.sleep(1)
    print(f"{asyncio.get_running_loop()=}")

if __name__ == "__main__":
    app()

[JosXa]

Is there any good reason that async commands are not part of the library? Or is it just a missing feature?

[borissmidt]

I really like typer that a function and command are the same thing. This allows me to reuse a command to chain bigger commands together. With argparse i have to manually write a mapping for each typed function i want to expose. Now we have to manually wrap each async function to become sync or subclass typer for each project we use typer in. While it would be better that this is handled by the typer library. Especially since asyncio is in the standard library. For example i could have an install dependency command which then reused the install dependency command for each dependency. With asyncio I could make this concurrently.

[takeda]

I can't see any appeal for having it at the moment other than it being extra code to be maintained to be honest. What's the problem with just calling aiorun?

Why not use argparse and not have extra dependencies? It's the same kind of argument. Ironically with how argparse there's less work to do this, because you don't have to place asyncio.run() for every subcommand. Given how typer operates and interest in this ticket I believe this is necessary.

[gpkc]

Of course none of those are fair comparisons. It's still FastAPI's sibling because it follows the same design philosophy with types and clarity. Not to mention that async support is much more needed on a web framework, which allows it to handle requests concurrently. For a CLI, it's not the case that you really need concurrency in most scenarios.

So, in the end, the appeal to support async on Typer is purely from a compatibility perspective. For example, you might have a FastAPI project and your code will have many async parts, it can be handy to be able to call those directly from your Typer CLI.

In any case, it was just a devil's advocate positioning, no need to get emotional over it :) I'm assuming these are probably the reasons why the support for async hasn't been done yet, and of course if it's actually denied everyone can just fork the project and add it too, then use as a dependency, so you don't have to repeat anything.

[gpkc]

I didn't mean compatibility with FastAPI. I meant compatibility with async projects' codebase in general.

If you're making multiple remote requests with HTTPX, you could organize your code to, for example, chain aiorun with asyncio.gather. If you're making a single request, there's no advantage on using asyncio as httpx supports synchronous calls.

Again, this is just some basic questioning which I assume the package maintainers will think about and are probably the reason why thins hasn't been fixed yet.

[takeda]

Of course none of those are fair comparisons. It's still FastAPI's sibling because it follows the same design philosophy with types and clarity. Not to mention that async support is much more needed on a web framework, which allows it to handle requests concurrently. For a CLI, it's not the case that you really need concurrency in most scenarios.

Just because you use async only for web doesn't mean everyone else is. There are many async-only libraries that have little to do with web.

[takeda]

Nobody would be complaining about missing async support if was just for compatibility (similarity actually) with FastAPI. There's a increasingly common case of calling async functions, IMO the main example would be remote requests with httpx. I must admit I'm really surprised with the resistance over the 15 lines of code this change represents.

@rafalkrupinski, that person has nothing to do with this project he is just there to waste everyone's time, because he only uses async when using "web framework" and nothing else.

[gpkc]

Just because you use async only for web doesn't mean everyone else is. There are many async-only libraries that have little to do with web.

Hmm you seem to lack context or interpretation. I wasn't even responding to just you, but also to someone else who was comparing Typer with FastAPI, hence my response was focused on the comparison between the two. But interesting that you are able to instantly infer my whole dev experience from this brief conversation.

This issue has been open for 3 years with no traction forward so I'm just trying to figure out why that is the case :) since I don't think there's been any official positioning about it. I'm pretty sure my questions won't change the maintainers' mind if they are leaning towards incorporating this into the package, when they get time to come here it should actually help if it doesn't get derailed by this kind of comment. Also, I've even helped formulate a solution for it that could be incorporated in the official implementation here...

In any case, just to address your "answer", it doesn't matter if there are other async-only libraries and/or if they have to do with web. The point is that adding async capabilities to Typer means there should be an appeal for it to handle concurrency. This means now Typer will have to deal with the async loop in some form. If this hasn't been added yet, maybe the maintainer's philosophy is that this isn't the case and that concurrency should be handled by the user.

[francipvb]

Hello all! Thanks for all the discussion and workarounds.

I want to add support for async in Typer, without making it required or default, using AnyIO as an optional dependency. In FastAPI/Starlette, AnyIO is required, as everything is just async underneath. But in Typer I don't want to force people to use AnyIO when they don't really need it. That makes the whole thing a bit more complex as I need to add all the logic to conditionally use or not AnyIO and do async or not.

This is one of the top priorities for Typer, along with a code reference (API reference). I'm finishing some things in FastAPI, I have some things in SQLModel too, and then I'll continue with Typer stuff.

Hello @tiangolo,

At least for me, by putting AnyIO as a dependency is not a problem. Actually, an async variant of the typer with specialyzed algorythm to run async functions could be enough.

In this case a check if AnyIO is importable could be put in the AsyncTyper constructor.

Thank you for these libraries you develop.

[JosXa]

@borissmidt's PR is still open. Not sure about 'hostility'

[hhstore]

Any news on the async support? Any way we can help?

with syncer:

• ✅ try this lib: https://github.com/miyakogi/syncer

pip install syncer

usage:

import os

import typer
from faststream.nats import NatsBroker
from loguru import logger
from syncer import sync

app = typer.Typer()

HOST = os.getenv("NATS_HOST", default="nats://localhost:4222")

async def aio_pub(host: str, msg: str, subject: str):
    async with NatsBroker(host) as broker:
        logger.debug(f"publish message: {msg}")
        await broker.publish(msg, subject=subject)

@app.command("pub")
def pub_task(host: str = HOST):
    #
    # here! converter: async -> sync
    #
    f_pub = sync(aio_pub)

    # sync call:
    f_pub(host, "Hi async wrap!", "test")

    logger.debug("pub task done")


if __name__ == "__main__":
    app()