Futures Concurrency I: Introduction
— 2019-09-02

  1. futures concurrency
  2. infallible: join
  3. infallible: select
  4. summary: infallible functions
  5. fallible: join
  6. fallible: select
  7. summary: fallible functions
  8. variadics
  9. methods
  10. async concurrency: javascript
  11. async concurrency: golang
  12. pattern: log all errors in try_select
  13. conclusion

In a previous post we've looked at Rust streams. In this post we're going to discuss another problem in the async space: futures concurrency combinators.

We're going to cover the different forms of concurrency that can be expressed with Futures, and cover both fallible and infallible variants.

For the sake of this article I'm going to assume you're already familiar with Futures as a mechanism for scheduling async work. If you're not, I recommend starting with the Rust async book.

Futures Concurrency

In general when you're running Futures you can schedule them in one of 3 ways:

The first way of scheduling futures is "sequential", and the default if we await multiple futures in a row. The other two scheduling approaches are "concurrent", and have some nuances and variations to them. We also distinguish between "fallible" and "infallible", where we'll start with the infallible versions first.

Infallible: Join

The first concurrent approach is "parallel". In this mode we have a multiple futures that we all want to complete. We await them all together, and write the result of each future to a tuple. An example of this:

use async_std::future;

let a = future::ready(1);
let b = future::ready(2);

let c = future::join(a, b);
assert_eq!(c.await, (1, 2));

Where the function signature would be:

pub async fn join<L, R, T1, T2>(left: L, right: R) -> (T1, T2)
    where L: Future<Output = T1>,
          R: Future<Output = T2>;

If we want to join more than two futures we can nest calls to join, and later in this post we'll look at possible ways to make this an even better experience.

Infallible: Select

The other form of concurrency is having multiple futures, and wanting one to resolve. This is generally less common than waiting for all futures to resolve, but still useful when for example wanting to set a time limit on a future.

There are many concurrency scenarios imaginable, but in general this means waiting for the first future to resolve. This is generally done through the select function. An example:

use async_std::future;

let a = future::pending();
let b = future::ready(1);

let c = future::select(a, b);
assert_eq!(c.await, 1);

Where the function signature would be:

pub async fn select<F, T>(left: F, right: F) -> T
    where F: Future<Output = T>;

Summary: Infallible functions

For operating on infallible futures (e.g. futures that don't return Result), we've defined the following methods:

NameReturn signatureWhen does it return?
future::join(T1, T2)Wait for all to complete
future::selectTReturn on first value

Let's move onto fallible futures next!

Fallible: Join

When calling join on two futures that return Result, being handed back a (Result<T, E>, Result<T, E>) makes error handling rather awkward:

use async_std::future;

let a = future::ready(Ok(1));
let b = future::ready(Ok(2));

let c = future::join(a, b).await;
assert_eq!((c.0?, c.1?), (1, 2)); // this is not great

Instead it'd be much nicer if we were handed back a Result<(T1, T2), E>, so we could handle errors more or less like we'd expect to:

use async_std::future;

let a = future::ready(Ok(1));
let b = future::ready(Ok(2));

let c = future::try_join(a, b);
assert_eq!(c.await?, (1, 2)); // much better

This introduces a new function try_join with a signature of:

pub async fn try_join<L, R, T1, T2, E>(left: L, right: R) -> Result<(T1, T2), E>
    where L: Future<Output = Result<T1, E>>,
          R: Future<Output = Result<T2, E>>;

An implication of the API returning Result<(T1, T2), E> is that if either of the futures returns Err, we should drop the other future because it will never yield its result. This is commonly referred to as early rejection.

If it's undesirable to reject early, using join instead of try_join allows all futures to complete before returning.

Fallible: Select

Like we said earlier, the purpose of select is to get back the first result from multiple futures. However, say we want to check both our local cache and a remote cache if something is stored there, we don't necessarily care if either of them fails, as long as one of them has the result we want. We cannot express this using the regular select function:

use async_std::future;

let a = future::ready(err!("oh no!")); // some error
let b = future::ready(Ok(1));

let c = future::select(a, b);
assert!(c.await.is_err()); // if `a` is an error, we cannot proceed to `b`

The semantics we want are: "give back the first response that succeeds, or if none of them succeed return the last error."

use async_std::future;

let a = future::ready(err!("oh no!")); // some error
let b = future::ready(Ok(1));

let c = future::try_select(a, b);
assert_eq!(c.await?, 1); // `a` failed, `b` succeeded -- so we get `b`

Where the function signature would be:

pub async fn try_select<F, T, E>(left: F, right: F) -> Result<T, E>
    where F: Future<Output = Result<T, E>>;

If it's undesirable to get keep trying until all options have been exhausted, then it's acceptable to use select for fallible futures too, to get the first future regardless of whether it succeeded, and reject early if it failed.

Summary: Fallible functions

When talking about fallible futures, we have access to a total of 4 functions:

NameReturn signatureWhen does it return?
future::join(Result<T, E>, Result<T, E>)Wait for all to complete
future::try_joinResult<(T1, T2), E>Return on first Err, wait for all to complete
future::selectResult<T, E>Return on first value
future::try_selectResult<T, E>Return on first Ok, reject on last Err

In most cases the try_ functions will be the right choice for fallible futures, but in the off chance it's not, the regular variants allow bypassing the semantics to do more custom things. And even better: by mixing both kinds of functions, more complicated flows can be created.

Variadics

The functions we've talked about so far are quite nice, but it's not rare to want join or select more than two futures. A way to do this is to apply combinators multiple times:

use async_std::future;

let a = future::ready(1);
let b = future::ready(2);
let c = future::ready(3);

let j = future::join(a, b);
let k = future::join(j, c);
assert_eq!(k.await, ((1, 2), 3)); // this is probably not great to work with though

Instead it'd be much nicer if we could provide an n length input, and get back an n length tuple. The way to do this is through variadics; functions with varying input. The way we can write variadics in Rust today is by using macros:

use async_std::future;

let a = future::ready(1);
let b = future::ready(2);
let c = future::ready(3);

let j = future::join!(a, b, c);
assert_eq!(j.await, (1, 2, 3)); // much better

Being able to accept multiple arguments is useful for all methods we've described so far. Additionally all of them could, and probably should, be written as variadics rather than regular functions.

Methods

So far we've only talked about "free functions" (standalone functions), and not "methods" (functions that are part of a struct). In Rust, structs can't have associated macros, which means that variadics wouldn't be possible there, which in turn means we'd lose out on ergonomics.

But there's another reason: to our knowledge it's currently impossible to define the try_ variants directly on Future without introducing an intermediate TryFuture type.

In the future specialization might make it possible to implement the try_ methods for futures with the right signature. But in the mean time if we're thinking of the core concurrency mechanisms and how they may some day live in std, having them be variadic free functions seems like the better choice all around.

Async Concurrency: JavaScript

JavaScript is an asynchronous-first language that in recent years has chosen Promise + async/await as its core concurrency mechanism. There are two core differences between JavaScript Promises and Rust's Futures:

  1. When a Promise is created it is immediately scheduled on the microtask queue (executor), while in Rust a Futures is only scheduled when .awaited.
  2. Every Promise in JavaScript is fallible (reject/resolve), while a Rust Future can be either fallible or infallible, depending on whether it returns a Result.

Since the last TC39 meeting, JavaScript has the following 4 methods for concurrency in Promises:

namesignaturedescriptionstage
Promise.allSettled(iterable)does not short-circuit4 (stable)
Promise.all(iterable)short-circuits when an input value is rejected4 (stable)
Promise.race(iterable)short-circuits when an input value is settled4 (stable)
Promise.any(iterable)short-circuits when an input value is fulfilled2 (experimental)

What we're proposing in this post is fairly similar, but unlike JavaScript we're able to discern between fallible / infallible Futures which makes allows us to make use of that in the design. In fact, for fallible Futures they compare really well:

JavaScriptRustdescription
Promise.allSettledfuture::joindoes not short-circuit
Promise.allfuture::try_joinshort-circuits when an input value is rejected
Promise.racefuture::selectshort-circuits when an input value is settled
Promise.anyfuture::try_selectshort-circuits when an input value is fulfilled

Async Concurrency: Golang

So I'm by no means a Golang expert, but I know we've been looking at it quite a bit when trying to figure out futures concurrency. In particular the Golang select blocks have an equivalent in the existing futures::select macro:

use futures::select;
let mut a = future::ready(4);
let mut b = future::pending::<()>();

let res = select! {
    a_res = a => a_res + 1,
    _ = b => 0,
};
assert_eq!(res, 5);

It seems well-understood that the macro in its current form wouldn't make it into the standard library. Which would mean that this form of select would probably need to be elevated to the language level, which would require a fair amount of design work, and at waiting until at least a new edition.

Side note: this is not to say that I'm proposing the addition of any API to stdlib quite yet. But I like to consider the possibility that at some point in the future we may want to. So in order to have a fruitful conversation about whether we should, I want to make sure we don't have to question whether we could.

While select blocks seem cool, I think it'd be preferable if we could solve the same problems without needing changes to the language. And I think with the select / try_select functions we're proposing we do!

The biggest difference with Go's select blocks is that all "branches" in our select functions resolve to the same type, while in at least the Rust/Go version it seems like it's possible to return multiple return types trough an Either type.

Pattern: log all errors in try_select

One question that comes up frequently is: "how do I log all errors inside a select call?" The way to do that is to map_err (or equivalent) the future before passing it to select:

let a = a.map_err(|e| dbg!(e));
let b = b.map_err(|e| dbg!(e));
let c = future::try_select(a, b).await?;

If map_err is not available (as is the case with async_std), it's possible to achieve the same behavior by defining a new async block:

let a = async { Ok(a.await.map_err(dbg!(e))?) }
let b = async { Ok(b.await.map_err(dbg!(e))?) }
let c = future::try_select(a, b).await?;

Conclusion

In this post we've covered the challenges of both fallible and infallible async concurrency in Rust, compared it to approaches from two different languages, and formulated a coherent model that covers all cases.

The next step for this proposal is to write experimental implementations for async-std to see how well it works in practice.

In the future we may also want to cover streams concurrency. In particular ways we could improve upon Stream::for_each_concurrent / Stream::try_for_each_concurrent / Stream::FuturesUnordered seem like an interesting topic. But that's not in scope for today.

Thanks all for reading, and have a good week!

Special thanks to Tirr-c and David Barsky for proof reading!