Can you expand more on why these AI cases make the complexity tradeoff different?
I'd imagine think waiting on a 3rd party LLM API call would be computationally very inexpensive compared to what's going on at the business end of that API call. Further lowering the cost, Django is usually configured to use multiple threads and/or processes so that this blocking call won't keep a CPU idle, no?
They’re very slow. Like, several seconds to get a response slow. If you’re serving a very large number of very fast requests, you can argue that the simplicity of the sync model makes it worth it to just scale up the number of processes required to serve that many requests, but the LLM calls are slow enough that it means you need to dramatically scale up the number of serving processes available if you’re going to keep the sync model, and that’s mostly to have CPUs sitting around idle waiting for the LLM to come back. The async model can also let you parallelize calls to the LLMs if you’re making multiple independent calls within the same request - this can cut multiple seconds off your response time.
If the alternative to Async is more processes, rather than threads, a clear benefit is reduced memory usage and reduced process startup time.
Not necessarily true. Many process-parallel Python environments support using fork(2) for parallelism (multiprocessing, gunicorn, celery).
For similar processes (e.g. parallel waiting on RPCs) that removes the memory overhead. It also largely mitigates startup time costs (especially if forks are reused for multiple requests, which they are in most forking contexts).
While there is debate and grumbling in the Python community about fork(2)’s rough edges re: signals/threads/MacOS, these issues are usually handled inside parallelism-management library code and rarely concern application level developers.
To restate my comment: I argued (1) this CPU cost would be very marginal compared to the LLM API compute cost, and (2) the CPU blocking claim doesn't really hold, due to the wonders of threads and processes.
1. A blocking call, which can take 3-10 seconds.
2. A streaming call, which can also take 3-10 seconds but where the content is streaming directly to you as it is generated (and you may be proxying it through to your user).
In both cases you risk blocking a thread or process for several seconds. That's the problem asyncio solves for you - it means you could have hundreds (or even thousands) of users all waiting for the response to that LLM call without needing hundreds or thousands of blocked threads/processes.
Or maybe the disconnect is where I'm taking for granted that "cost of parked thread" is the same as worrying about the nr of parked threads? Maybe everyone uses Django setups where it's nontrivial to add memory, increase serverless platform limits, etc if you get the happy problem of 10k concurrent users? Or maybe people don't know that the nr of threads/processes you can have on Linux is much more than hundreds or thousands. Or maybe there's some Python or Django specific limits to this.
See eg https://www.baeldung.com/linux/max-threads-per-process for some runtime sysctl knobs if you want to go higher than the default limits. (Though the section 6 there seems out of date and written with the 32-bit OS in mind)
[1] as reported by the Python snippet from https://stackoverflow.com/a/64406494