https://theta.eu.org/2021/03/08/async-rust-2.html e (eta) all posts source code Why asynchronous Rust doesn't work 2021-03-08 * permalink In 2017, I said that "asynchronous Rust programming is a disaster and a mess". In 2021 a lot more of the Rust ecosystem has become asynchronous - such that it might be appropriate to just say that Rust programming is now a disaster and a mess. As someone who used to really love Rust, this makes me quite sad. I've had a think about this, and I'm going to attempt to explain how we got here. Many people have explained the problems with asynchronous programming - the famous what colour is your function essay, for example.^1 However, I think there are a number of things specific to the design of Rust that make asynchronous Rust particularly messy, on top of the problems inherent to doing any sort of asynchronous programming. In particular, I actually think the design of Rust is almost fundamentally incompatible with a lot of asynchronous paradigms. It's not that the people designing async were incompetent or bad at their jobs - they actually did a surprisingly good job given the circumstances!^2 I just don't think it was ever going to work out cleanly - and to see why, you're going to have to read a somewhat long blog post! A study in async I'd like to make a simple function that does some work in the background, and lets us know when it's done by running another function with the results of said background work. use std::thread; /// Does some strenuous work "asynchronously", and calls `func` with the /// result of the work when done. fn do_work_and_then(func: fn(i32)) { thread::spawn(move || { // Figuring out the meaning of life... thread::sleep_ms(1000); // gee, this takes time to do... // ah, that's it! let result: i32 = 42; // let's call the `func` and tell it the good news... func(result) }); } There's this idea called "first-class functions" which says you can pass around functions as if they were objects. This would be great to have in Rust, right? See that func: fn(i32)? fn(i32) is the type of a function that takes in one singular i32 and returns nothing. Thanks to first-class functions, I can pass a function to do_work_and_then specifying what should happen next after I'm done with my work - like this: fn main() { do_work_and_then(|meaning_of_life| { println!("oh man, I found it: {}", meaning_of_life); }); // do other stuff thread::sleep_ms(2000); } Because do_work_and_then is asynchronous, it returns immediately and does its thing in the background, so the control flow of main isn't disrupted. I could do some other form of work, which would be nice (but here I just wait for 2 seconds, because there's nothing better to do). Meanwhile, when we do figure out the meaning of life, it gets printed out. Indeed, if you run this program, you get: oh man, I found it: 42 This is really exciting; we could build whole web servers and network stuff and whatever out of this! Let's try a more advanced example: I have a database I want to store the meaning of life in when I find it, and then I can run a web server in the foreground that enables people to get it once I'm done (and returns some error if I'm not done yet). struct Database { data: Vec } impl Database { fn store(&mut self, data: i32) { self.data.push(data); } } fn main() { let mut db = Database { data: vec![] }; do_work_and_then(|meaning_of_life| { println!("oh man, I found it: {}", meaning_of_life); db.store(meaning_of_life); }); // I'd read from `db` here if I really were making a web server. // But that's beside the point, so I'm not going to. // (also `db` would have to be wrapped in an `Arc>`) thread::sleep_ms(2000); } Let's run this...oh. error[E0308]: mismatched types --> src/main.rs:27:22 | 27 | do_work_and_then(|meaning_of_life| { | ______________________^ 28 | | println!("oh man, I found it: {}", meaning_of_life); 29 | | db.store(meaning_of_life); 30 | | }); | |_____^ expected fn pointer, found closure | = note: expected fn pointer `fn(i32)` found closure `[closure@src/main.rs:27:22: 30:6]` I see. Hang on a minute... So, this is actually quite complicated. Before, the function we passed to do_work_and_then was pure: it didn't have any associated data, so you could just pass it around as a function pointer (fn (i32)) and all was grand. However, this new function in that last example is a closure: a function object with a bit of data (a &mut Database) tacked onto it. Closures are kind of magic. We can't actually name their type - as seen above, the Rust compiler called it a [closure@src/main.rs:27:22: 30:6], but we can't actually write that in valid Rust code. If you were to write it out explicitly, a closure would look something like this: struct Closure<'a> { data: &'a mut Database, func: fn(i32, &mut Database) } impl<'a> Closure<'a> { fn call(&mut self, arg: i32) { (self.func)(arg, self.data) } } There are a number of things to unpack here. --------------------------------------------------------------------- An aside on naming types Being able to name types in Rust is quite important. With a regular old type, like u8, life is easy. I can write a function fn add_one (in: u8) -> u8 that takes one and returns one without any hassle. If you can't actually name a type, working with it becomes somewhat cumbersome. What you end up having to do instead is refer to it using generics - for example, closures' types can't be named directly, but since they implement one of the Fn family of traits, I can write functions like: fn closure_accepting_function(func: F) where F: Fn(i32), // <-- look! { /* do stuff */ } If I want to store them in a struct or something, I'll also need to do this dance with the where clause every time they're used. This is annoying and makes things harder for me, but it's still vaguely workable. For now. msql-srv example code [image: from the msql-srv crate, showing an example of many where clauses as a result of using closures] --------------------------------------------------------------------- An aside on 'radioactive' types The way Rust is designed tends to encourage certain patterns while discouraging others. Because of ownership and lifetimes, having pieces of data that hold references to other pieces of data becomes a bit of a problem. If my type has a & or a &mut reference to something, Rust makes me ensure that: * the something in question outlives my type; you can't go and drop the thing I'm referring to if I still have a reference to it, otherwise my reference will become invalid * the something in question doesn't move while I have the reference * my reference to the something doesn't conflict with other references to the something (e.g. I can't have my & reference if something else has a &mut reference) So types with references in them are almost 'radioactive'; you can keep them around for a bit (e.g. inside one particular function), but attempting to make them long-lived is usually a bit of an issue (requiring advanced tricks such as the Pin type which didn't even exist until a few Rust versions ago). Generally Rust doesn't really like it when you use radioactive types for too long - they make the borrow checker uneasy, because you're borrowing something for an extended period of time. borrowing semantics visualization [image: from the RustViz paper, showing borrowing semantics] --------------------------------------------------------------------- Closures can be pretty radioactive. Look at the thing we just wrote out: it has a &'a mut Database reference in it! That means while we're passing our Closure object around, we have to be mindful of the three rules (outlives, doesn't move, no conflicting) - which makes things pretty hard. I can't just hand off the Closure to another function (for example, the do_work_and_then function), because then I have to make all of those rules work, and that's not necessarily easy all the time. (Not all closures are radioactive: if you make them move closures, they'll take everything by value instead, and create closure objects that own data instead of having radioactive references to data. Slightly more of a pain to deal with, but you lose the blue radiation glow the things give out when you look at them.) Also, remember what I said about being able to name types? We're not actually dealing with a nice, written-out Closure object here; we're dealing with something the compiler generated for us that we can't name, which is annoying. I also lied when I said that it was as simple as making all of your functions take F, where F: Fn(i32) or something - there are actually three different Fn-style traits, Fn, FnMut, and FnOnce. Do you know the difference between them? So. A closure is this magical, un-nameable type that the compiler makes for us whenever we use || {...} syntax, which implements one of three traits (and it's not immediately obvious which), and it also might be radioactive. Try and use one of these, and the Rust compiler is probably going to be watching you very carefully. --------------------------------------------------------------------- The thing I really want to try and get across here is that Rust is not a language where first-class functions are ergonomic. It's a lot easier to make some data (a struct) with some functions attached (methods) than it is to make some functions with some data attached (closures). Trying to use ordinary structs is downright easy: * they're explicitly written out by the programmer with no funky business * you choose what traits and methods to implement on them and how to set them out / implement them * the struct can actually be referred to by other parts of the code by its type Trying to use closures is hard: * the compiler does some magic to make a closure type for you * it implements some obscure Fn trait (and it's not immediately obvious which) * it might be radioactive (or force you to use move and maybe insert a bunch of clone() calls)^3 * you can't actually name their type anywhere or do things like return them from a function Importantly, the restrictions applied to using closures infect types that contain them - if you're writing a type that contains a closure, you'll have to make it generic over some Fn-trait-implementing type parameter, and it's going to be impossible for people to name your type as a result. (Other languages, like Haskell, flip this upside down: functions are everywhere, you can pass them around with reckless abandon, etc. Of course, these other languages usually have garbage collection to make it all work...) --------------------------------------------------------------------- Bearing this in mind, it is really quite hard to make a lot of asynchronous paradigms (like async/await) work well in Rust. As the what colour is your function post says, async/await (as well as things like promises, callbacks, and futures) are really a big abstraction over continuation-passing style - an idea closely related to the Scheme programming language. Basically, the idea is you take your normal, garden-variety function and smear it out into a bunch of closures. (Well, not quite. You can read the blue links for more; I'm not going to explain CPS here for the sake of brevity.) Hopefully by now you can see that making a bunch of closures is really not going to be a good idea (!) *wibbly wobbly scene transition* And then fast forward a few years and you have an entire language ecosystem built on top of the idea of making these Future objects that actually have a load of closures inside^4, and all of the problems listed above (hard to name, can contain references which make them radioactive, usually require using a generic where clause, etc) apply to them because of how "infectious" closures are. The language people have actually been hard at work to solve some (some!) of these problems by introducing features like impl Trait and async fn that make dealing with these not immediately totally terrible, but trying to use other language features (like traits) soon makes it clear that the problems aren't really gone; just hidden. Oh, and all the problems from what colour is your function are still there too, by the way - on top of the Rust-specific ones. Beginner (and experienced) Rust programmers look at the state of the world as it is and try and build things on top of these shaky abstractions, and end up running into obscure compiler errors, and using hacks like the async_trait crate to glue things together, and end up with projects that depend on like 3 different versions of tokio and futures (perhaps some async-std in there if you're feeling spicy) because people have differing opinions on how to try and avoid the fundamentally unavoidable problems, and it's all a bit frustrating, and ultimately, all a bit sad. --------------------------------------------------------------------- Did it really have to end this way? Was spinning up a bunch of OS threads not an acceptable solution for the majority of situations? Could we have explored solutions more like Go, where a language-provided runtime makes blocking more of an acceptable thing to do? Maybe we could just have kept Rust as it was circa 2016, and let the crazy non-blocking folks^5 write hand-crafted epoll() loops like they do in C++. I honestly don't know, and think it's a difficult problem to solve. But as far as my money goes, I'm finding it difficult to justify starting new projects in Rust when the ecosystem is like this. And, as I said at the start, that makes me kinda sad, because I do actually like Rust. --------------------------------------------------------------------- (Common Lisp is pretty nice, though. We have crazy macros and parentheses and a language ecosystem that is older than I am and isn't showing any signs of changing...) --------------------------------------------------------------------- 1. This is really recommended reading if you aren't already familiar with it (as you'll soon see) - 2. Seriously - when I put out the last blog post, the actual async core team members commented saying how much they appreciated the feedback, and then they actually went and made futures 1.0 better as a result. Kudos! - 3. You might be thinking "well, why don't you just only use move closures then?" - but that's beside the point; it's often a lot harder to do so, because now you might have to wrap your data in an Arc or something, by which point the ergonomic gains of using the closure are outweighed by the borrow checker-induced pain. - 4. You can actually manually implement Future on a regular old struct. If you do this, things suddenly become a lot simpler, but also you can't easily perform more async operations inside that struct's methods. - 5. (sorry, I mean, the esteemed companies that deign to use Rust for their low-latency production services) - subscribe here (RSS) or here (email), if you'd like to be notified of new posts This person's ramblings are distributed in the hope that they will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. (university lecture notes!) * (site archives) * (blah)