Asynchronous Server Gateway Interface

And for instance, you know, your, your blog application.

Well, you know, how d- how does it go running under gunicorn with, you know, a, a pre-forked worker versus a threaded worker versus, well, what about if you run it as a, as a, as an ASCII application, and how does that perform at the sc- at the scales that you're interested in? So we also b- you know, it also load tests the application, and you're like, "Well, you know, it, it might fall over at 500 requests per second, but, you know, I'm not getting 500 requests per second to my blog.

So do you know what? My, my gunicorn worker with its two threaded workers is perfectly fine." Um, the, the, the idea is to be able to, um, talk a little bit more sensibly about, um, how you should deploy your Django application than just like, "Oh, look, it does so many requests per second," which is a little bit, you know, vacuous, basically.

It's like, well, n- you know, I want to deploy a wagtail app.

Which brings us to the third component, the server application itself.

The training modules specify building this application using Python web frameworks, specifically mentioning Starlette, and combining it with an ASGI server like Uvicorn to actually listen on the network port.

If we connect this to the bigger picture, this is a masterclass in lowering the barrier to entry.

Microsoft isn't asking developers to learn a brand-new proprietary networking protocol just for AI.

Exactly.

Starlette is a wildly popular lightweight web framework.

But the mention of ASGI, which stands for Asynchronous Server Gateway Interface, that is the technical linchpin here.

Why is an asynchronous server so important for AI specifically?

Yeah, no problem.

What about if we put it under load? Or what about if we're running ASCII under load? What about if we're running with free-threaded Python? How does that change the thing? So to be able to construct these different scenarios, run them with Locus to show their actual performance profile, I think that's going to do a lot more to sort of enable us to have a more sophisticated discussion about performance than, oh, look how many requests per second it does in this hypothetical benchmark.

Right, it's also

abstract,