Rust 2018 has shipped, and we’re closing in on the end of the year. While we didn’t manage to ship async/await as part of the edition itself, the community has made quite a lot of progress toward that goal. This post summarizes the state of play, and announces the publication of several crates intended to facilitate use of async/await on the nightly ecosystem.

Why async/await

Before delving into the current status, it’s worth taking a moment to recap the core motivations for async/await, and its special importance for Rust.

Async/await notation is a way of making asynchronous programming more closely resemble synchronous programming. To see how this works, consider Read::read in std::io:

fn read(&mut self, buf: &mut [u8]) -> Result<usize, std::io::Error>

This synchronous method blocks the current thread until data has been read into buf, then says how many bytes were read. We can build on this method to implement read_exact, a method that continues reading until the buffer is filled:

fn read_exact<T: Read>(input: &mut T, buf: &mut [u8]) -> Result<(), std::io::Error> { let mut cursor = 0; while cursor < buf.len() { cursor += input.read(&mut buf[cursor..])?; }
}

In the asynchronous world, we want to perform similar operations, but rather than blocking the current thread we want to leave it free to do other work while the I/O operations complete asychronously. But actually programming directly in that way is incredibly difficult. What we want is to program as if I/O operations will block the current thread, but have the compiler transform this code into more efficient asynchronous execution.

In short, our goal is to be able to write the following:

async fn read_exact<T: AsyncRead>(input: &mut T, buf: &mut [u8]) -> Result<(), std::io::Error> { let mut cursor = 0; while cursor < buf.len() { cursor += await!(input.read(&mut buf[cursor..]))?; }
}

Comparing the two snippets, there are three changes here:

  • We write async before fn, to signal that the function should be asynchronously executed on its parent thread. Async functions return their result within a Future, representing a value that must be asynchronously computed.

  • We use the AsyncRead trait (from futures::io), rather than the Read trait. This makes an asynchronous version of the read method available (currently provided via the AsyncReadExt extension trait).

  • We enclose the call to read with await!, signaling that we want to simulate blocking on the operation to complete.

And that’s all.

This approach to asynchrony has proven itself in many other languages already. But there’s an extra element in the Rust version: borrowing. For read_exact, we are able to hold the borrows of input and buf while we use await! (which may actually clear the stack and run completely unrelated code). The validity of the borrowing is still checked, and it works largely similarly to borrowing within synchronous code. (The main difference: the way elision works in fn signatures.)

If you’ve programmed with futures in Rust before, you’ll know this is a game changer: manual futures code generally must be restricted to 'static data, which (in addition to its verbosity) takes it far away from idiomatic Rust, forcing you to program with Arc and Mutex far more frequently than usual.

Below, we’ll check in on the status of various aspects of the transition to this new world.

The book

Part of the work around async/await this year has been writing a new book, covering the syntax, the underlying Future API, and ultimately various programming patterns that emerge. @cramertj has written an early draft (repo here), which is already useful for understanding these concepts.

The syntax

The async/await syntax itself has had an implementation on nightly for several months now, and is being used at a fairly large scale in Google’s Fuchsia project. You can find more detail about that usage here.

While there are a few remaining limitations in the implementation of the syntax, the main issue that remains to be resolved prior to any stabilization is the await side of the syntax. @withoutboats recently wrote a blog post describing the issues there in detail.

Today, async can be used in code blocks, for free functions (async fn), and for inherent methods. Ultimately it will be usable in trait method signatures as well, but this is effectively blocked by existential types, another feature on track to stabilization relatively soon. In the meantime, there are several forward-compatible ways to continue using async blocks within trait implementations, most simply by placing the async block into a Box that will be removable later, once existential types are stable.

The Supporting APIs

Like many other language features, async/await also requires some support in the standard library: shipping the Future trait (and associated machinery) in std. That work has been a major thrust this year, and is nearing completion.

Core futures APIs

There’s currently an open RFC proposing stabilization of the futures APIs, and includes a fairly detailed writeup of the history of those APIs. The pull request contains a checklist of current blockers; the most signifcant one at the moment is finalizing the Waker APIs.

The Pin API

One of the underlying mechanisms supporting the futures API is the Pin type, which is also how we enable borrowing in async blocks. This API, too, has seen significant iteration over the course of the year. @withoutboats’s blog post from a few months ago covers the final design, which has also been proposed for stabilization. The only remaining sticking point is around type and trait naming.

Compatibility with futures 0.1

The design of the futures API had to change in breaking ways in order to support async/await. However, there’s a large existing ecosystem of code that uses the earlier futures 0.1 API. Luckily, we’re able to provide a rather ergonomic compatibility layer that makes it possible to move between the two APIs easily, and hence support incremental migration. A recent blog post from @jsdw does an excellent job of laying out how this compatibility story works.

Some new crates

In addition to the compatibility layer the Networking Working Group has also put effort into building crates directly using the new futures API, in order to more fully vet that API, to provide a smoother experience for others wanting to build code using async/await, and to lay out a clear vision of what the new ecosystem might look like.

  • Romio, a minimal fork of Tokio based directly on the new futures API. While Tokio proper aims to provide a comprehensive and opinionated story for the lowest-levels of async networking code, Romio covers just the essentials: an API surface very similar to std::net, but supporting async/await directly. The crate includes a good bit of documentation and examples, and @withoutboats has written a blog post detailing lessons learned through this port.
  • http-service, a tiny crate building on bytes, http, and the new futures API to provide a common interface for http-based services using the new futures API. This crate is partly based on the ongoing work on Tide, where the goal is to seed the ecosystem with numerous small, useful crates of this kind that many different frameworks and libraries can build on. As such, the API is an extraction of the one initially used internally in Tide.
  • Tyger (forthcoming), a small crate that builds on top of Hyper to provide a direct http-service interface (and thus usable with async/await directly, without shims). Ultimately Tyger is likely to grow some other higher-level amenities, to complement Hyper’s relatively low-level focus. As with http-service, the crate is an early extraction from the Tide work and is intended to provide a small, community-driven building block that can be used by many other crates. It will be published some time in the next few weeks.

We’ve come a long way toward async/await in 2018! With the futures and pin APIs on the cusp of stabilization, we should very soon be in a position to propose stabilization of async/await proper, hopefully shipping in the first half of 2019. It will be crucial to continue to build out the library ecosystem around these APIs in the coming year. If you want to get involved in any of this exciting work, please drop by the “WG-Net” channels on the Rust Discord!