you're coming from a single-threaded perspective, where Twisted's partially-automated CPS transform is in effect emulating green threads under a cooperative sch

you're coming from a single-threaded perspective, where Twisted's partially-automated CPS transform is in effect emulating green threads under a cooperative scheduling model, where the "yield" represents yielding control of the current "thread" to potentially other "threads". Your concern over fully automating the transform is because it may introduce further cooperative threading "yield"s that you cannot control. Yes, that's exactly the case. To scale to multiple CPUs, I use distinct processes, not threads. Shared memory and preemptive execution require a lot of programming effort to play nicely together that I don't care to expend. The GIL in python means that threads generally won't ever let you use more than a single CPU anyway.