[904baf11]

> ASGI support

> Django 3.0 begins our journey to making Django fully async-capable by providing support for running as an ASGI application.

Bleh. ASGI is probably the worst implementation I can think of for a common async interface. Stringly-typed callbacks? async generators exist for a good reason. It's also very asyncio-centric, and asyncio is Not Good.

[florimondmanca]

I don't often comment on HN, but I feel like this time I have to step up.

Having worked and developed with it extensively for over a year now, ASGI is one of the most elegant and well-designed interfaces I've seen.

Just wanted to address this:

> async generators exist for a good reason.

While the idea of using async generators might sound attractive at first, there's so much more to ASGI than exchanging messages in both directions.

Even on that aspect, I'd be interested in knowing how you'd achieve bidirectional communication async generators. How would the ASGI server call into the application and handle messages coming both ways? Surely, it can't just `async for message in agen: ...`, as that would only replicate the `send()` ASGI callback — what about `receive()`?

Besides, how would middleware look? Would it be as simple as wrapping `receive()` and `send()` into whichever other callable adds the functionality your middleware needs?

And you probably wouldn't be able to support all the possible syntactical implementations allowed by the notion of a Python callable: functions, classes, callable class instances.

So, yeah, I think ASGI is a beast.

Also, I wholeheartedly congratulate the Django team for pushing the standard out there and driving the Python async web ecosystem forward. You rock!

[904baf11]

> Even on that aspect, I'd be interested in knowing how you'd achieve bidirectional communication async generators. How would the ASGI server call into the application and handle messages coming both ways? Surely, it can't just `async for message in agen: ...`, as that would only replicate the `send()` ASGI callback — what about `receive()`?

Yield is two-way. ``result = await agen.asend(next_message)` ` et al.

> Having worked and developed with it extensively for over a year now, ASGI is one of the most elegant and well-designed interfaces I've seen.

Having worked with Python async for 3 years, anything that asyncio touches is completely poisoned. ASGI is another in the line of terribleness of the modern async ecosystem.

> Besides, how would middleware look?

A second async generator.

> And you probably wouldn't be able to support all the possible syntactical implementations allowed by the notion of a Python callable

Require an async generator. Done!

> Also, I wholeheartedly congratulate the Django team for pushing the standard out there and driving the Python async web ecosystem forward. You rock!

Much like asyncio, something isn't good just because it's being pushed.

[takeda]

> Yield is two-way. `result = await agen.asend(next_message)` et al.

It's still limited. To get a result you need to send something. Using it for something that's not strictly request-response can get complex.

> Having worked with Python async for 3 years, anything that asyncio touches is completely poisoned. ASGI is another in the line of terribleness of the modern async ecosystem.

That's your own bias. I also used asyncio from the beginning and got a different opinion. The asyncio package that comes with python is low level and is meant for building frameworks. If you used aiohttp, asyncpg and other packages packages build on top of asyncio it's actually quite enjoyable.

[andrewgodwin]

I'd love to hear your suggestions for changes we could make while keeping it somewhat WSGI-compatible. It took a few years to refine it to where it is now, so it's not like we just threw something at the wall.

[jxcl]

It's much easier to say stuff sucks than to actually get a huge project like this shipped. Thanks so much for your hard work Andrew + team!

[904baf11]

> while keeping it somewhat WSGI-compatible

There's the problem. WSGI is fundamentally flawed too - it could also be using generators for a two-way communication channel instead of stringly typed callbacks.

In a world where Python has optional static type hints, it would be nice to have concrete objects passed too.

[orf]

Can you expand on the stringly typed callbacks? Where are they required? Are you referring to string keys in a callback dictionary, in which case stringly-typed seems a bit of an odd choice of words.

[904baf11]

Calling the receive async function gets you a dictionary with a string type key, which defines the layout of the dictionary. It's not nice.

[Ralfp]

Can you explain why asyncio is not good?

[904baf11]

It's convuluted, it's not structured concurrency (https://vorpus.org/blog/notes-on-structured-concurrency-or-g...), fundamentally despite being async/await a lot of networking code is based on callbacks anyway (see: asyncio.Protocol), and I've had a lot of trouble debugging code that spawn tasks because asyncio code has zero obligation to care about any of the tasks it spawns.

[tomchristie]

Yo. Two points here.

* You can absolutely stay within structured concurrency constraints using ASGI.

* The ASGI spec doesn’t have anything to do with asyncio. It’s purely an async/await interface - You can build against it just fine using other concurrency backends.