anyio support for hikari

[A5rocks]

These thoughts have been swirling around my head for the past few days and while I'm sure this shouldn't be added, I feel it deserves being written down.

Please read the whole thing and don't post a knee-jerk reaction, kthx.

anyio?

anyio is a wrapper library over either trio or asyncio, offering a interface for structured concurrency.

Why?

Well, I really really like SC :)

Beyond that, I believe it composes itself MUCH more nicely than, say, asyncio. Here's a somewhat biased example to show some concepts:

# currently:
message = await event_manager.wait_for(MessageCreateEvent, timeout=120)
# with SC
message = None
with anyio.move_on_after(120):
  async with event_manager.stream(MessageCreateEvent) as stream:
    message = await stream.next()

"But hey! That's more code!"

In my view, it's more composable concepts:

1. an anyio style API would probably emphasize stream more, making the stream part decently obvious (in fact dispatching would probably use streams under the hood but I'm not sure)
2. anyio.move_on_after or anyio.fail_after can be applied anywhere. If someone wants to make anything have a timeout, this will work on it -- unless it uses asyncio.spawn :-)
3. await stream.next() is a bit annoying, but oh well.

Another possibly more convincing reason is that asyncio.TaskGroup is planned (AFAIK) which will bring some aspects of SC to hikari, so some design changes would be nice nonetheless.

One final reason to use anyio is that it allows people to use alternative event loops (currently only trio) with hikari.

Why not?

I can think of two major reasons: dependencies and performance.

The dependency problem

hikari relies on aiohttp which is asyncio specific. Well, ouch.

It's used for

• HTTP client: httpx exists and in fact is probably better than aiohttp's HTTP client -- httpx supports HTTP/2 behind an extra!
• Websocket client: this is the real iffy part. There's some websocket implementations for anyio, but IMO based on the few I've looked at they don't meet all of hikari's requirements. I propose that instead, httpcore is used to make the connection and then wsproto is wired up such that there is a websocket.

The performance problem

anyio is basically an indirection layer; it's an extra function call at best which can cause extra overhead. This does not seem to be too big of a problem though as starlette's performance only dropped by 4.5% when doing a switch to anyio.

The bigger problem is the websocket implementation. I don't have any examples to point to, but I believe aiohttp to be highly optimized unlike whatever hikari would throw together. I propose that this be made and benchmarked first if anything is to be done, as I believe this is the main point that determines whether this is OK.

Restructuring

I believe some parts of hikari have to be changed to fit a SC model where you can only spawn in a TaskGroup/Nursery. I'm not familiar enough with the code but essentially anywhere using asyncio.spawn would probably have to be rethought.

An observation I've noted is that SC prefers memory channels over spawning tasks. I'm not sure what the perf impact is, but memory channels can be turned into an API that spawns tasks quite easily:

async with anyio.create_task_group() as tg: 
  async with event_manager.stream(object) as stream:
    async for event in stream:
      for listener in self._lookup_listeners(event):
        tg.spawn(listener, event)

Is this a complete document?

No way! I'm certain I'm missing quite a few points for and against and I think I'll be able to answer questions.

Also, there may be some deal-breaking thing that means this support cannot be done. If so, oh well I'm writing my own library anyways.

[davfsa]

Just realized that I never actually gave my opinion here

I personally think that this would be a nice addition to hikari, with the unfortunate need for a (partial) rewrite. For the foreseeable future, I don't think this is something I will be looking to try and implement, as I rather focus on stabilizing what we have and building on top of it.

This does not mean that it might not get done, I do find the idea quite interesting and it is now in my radar of things to do if I ever get bored.

Thanks for bring it to my attention (and sorry for taking so long to write something, just noticed now that I never answered)!