This can even be extended to "awaiting across function calls" with effect polymorphism. I haven't seen this done in this context, but it would allow things like passing an async closure, or a generic type with an async trait implementation, to a normal function and having it automatically become an async function instead.
The real choice between explicit and implicit suspension points is thus syntactic, not implementation-driven. The reasoning there is more along the lines of "I like to see where my function might suspend" vs "I don't want to pepper my code with `await`."
I've done that a fair bit in Scala using HKT. E.g. superclass is written in terms of a generic monadic type, one subclass implementation uses Future (actually EitherT with Future), another uses identity, a third uses Writer to track statistics...
The difference is not that continuations allow anything at all, but that your effects are commutative rather than layered like a monad transformer stack.
Fair enough, but you're effectively talking about a secondary, novel type system, right? One of the things I like about monads is that they can be just plain old values of plain old datatypes.
> The difference is not that continuations allow anything at all, but that your effects are commutative rather than layered like a monad transformer stack.
Yeah, monad transformers do feel overly cumbersome for the cases where effects do commute. The trouble is that some effects don't commute, and I've yet to see a formalism that had a good solution to distinguishing effects that commute from those that don't. (I read a paper about a tensor product construct once, but couldn't really follow how it was supposed to work in practice)
Yes, exactly.
The main reason I prefer language-level effects rather than monads is for composability with normal imperative control flow. Monads are written in terms of higher order functions passing closures around, and that gets really messy in a language where you not only have early return and loops with break and continue, but you also have destructors that need to run as you enter and leave scopes.
You can add that kind of stuff as monads but it gets really messy, and is basically completely untenable in a language like Rust that also cares about memory management. Even if Rust did have HKT, it would still be impossible to write a Monad trait that supports them all, for example.
This article makes some great points on the subject: http://blog.paralleluniverse.co/2015/08/07/scoped-continuati...
The first thing to say on this topic is that virtually all of this (including async/await) is a library in Rust, not a part of the language. You can build whatever model you want on top. The one I'm describing is the futures/tokio model, which is becoming the de-facto, default one. But if you want something else, you can do it! This already is a divergence from Go, where you have what the runtime gives you, and that's it.
Fundamentally, Go's model is cooperative coroutines, where the runtime does the pre-emption for you. (You can also call it yourself but as you note, this isn't usual.) It'll do this when you request something from the network, when you sleep, or when you use a channel.
Tokio's model, on the other hand, is single-threaded. Instead of a ton of gorutines, there's an event loop, and you place a chain of futures onto it. Each chain of futures can sort of be thought of as a goroutine, but it's also different. For example, Rust can know at compile time exactly how big each futures' stack is, and so can do a single, exactly correct allocation. Anyway, each future in the chain handles yielding for you.
Async/await, in Rust, is about making those chains easier to write.
Anyway, I hope that helps. Maybe someone else will have a good comparison too. There's a lot of options in this space, and similar terminology that means the same, but slightly different things, so it can be hard to get your head around.