> Regarding pgbouncer - yes, it helps with connection pooling, but you still have the fundamental issue of 25 workers = 25 max concurrent LLM streams.

I guess my point is why are you picking an arbitrarily low number like 25? If you know that workers are going to be "waiting for tokens" most of the time, why not bump that number way, WAY up?

And I guess I should clarify - I'm coming into this outside of the Python space (I touch python because it's hard to avoid when doing AI work right now, but it's hardly my favorite language). Basically - having done a lot of GoLang, which uses goroutines in basically the same way Ruby uses Fibers (lightweight runtime managed thread replacements) I'll tell you up front - The orchestration level still matters a LOT, and you're going to be dealing with a lot of complexity there to make things work, even if it does mean that some lower level code can remain unaware (colorless).

Even good ol' fashioned c++ has had this concept bouncing around for a long time ( https://github.com/boostorg/fiber ). It's good at some things, but it's absolutely not the silver bullet I feel like you're trying to pitch it as here.