One interesting question about async fn in traits is whether or not we should label the trait itself as async. Until recently, I didn’t see any need for that. But as we discussed the question of how to enable “maybe async” code, we realized that there would be some advantages to distinguishing “async traits” (which could contain async functions) from sync traits (which could not). However, as I’ve thought about the idea more, I’m more and more of the mind that we should not take this step — at least not now.

After my last post on dyn async traits, some folks pointed out that I was overlooking a seemingly obvious possibility. Why not have the choice of how to manage the future be made at the call site? It’s true, I had largely dismissed that alternative, but it’s worth consideration. This post is going to explore what it would take to get call-site-based dispatch working, and what the ergonomics might look like.

Re-reading my previous post, I felt I should clarify why I called it the “soul of Rust”. The soul of Rust, to my mind, is definitely not being explicit about allocation. Rather, it’s about the struggle between a few key values — especially productivity and versatility1 in tension with transparency. Rust’s goal has always been to feel like a high-level but with the performance and control of a low-level one. Oftentimes, we are able to find a “third way” that removes the tradeoff, solving both goals pretty well.

In the last few months, Tyler Mandry and I have been circulating a “User’s Guide from the Future” that describes our current proposed design for async functions in traits. In this blog post, I want to deep dive on one aspect of that proposal: how to handle dynamic dispatch. My goal here is to explore the space a bit and also to address one particularly tricky topic: how explicit do we have to be about the possibility of allocation?

Last Friday, tmandry, cramertj, and I had an exciting conversation. We were talking about the design for combining async functions in traits with dyn Trait that tmandry and I had presented to the lang team on Friday. cramertj had an insightful twist to offer on that design, and I want to talk about it here. Keep in mind that this is a piece of “hot off the presses”, in-progress design and hence may easily go nowhere – but at the same time, I’m pretty excited about it.

Hi all! Welcome to 2022! Towards the end of last year, Tyler Mandry and I were doing a lot of iteration around supporting “dyn async trait” – i.e., making traits that use async fn dyn safe – and we’re starting to feel pretty good about our design. This is the start of several blog posts talking about where we’re at. In this first post, I’m going to reiterate our goals and give a high-level outline of the design.

A quick update to my last post: first, a better way to do what I was trying to do, and second, a sketch of the crate I’d like to see for experimental purposes. An easier way to roll our own boxed dyn traits In the previous post I covered how you could create vtables and pair the up with a data pointer to kind of “roll your own dyn”. After I published the post, though, dtolnay sent me this Rust playground link to show me a much better approach, one based on the erased-serde crate.

If you’re willing to use nightly, you can already model async functions in traits by using GATs and impl Trait — this is what the Embassy async runtime does, and it’s also what the real-async-trait crate does. One shortcoming, though, is that your trait doesn’t support dynamic dispatch. In the previous posts of this series, I have been exploring some of the reasons for that limitation, and what kind of primitive capabilities need to be exposed in the language to overcome it.

In the previous post, I talked about how we could write our own impl Iterator for dyn Iterator by adding a few primitives. In this post, I want to look at what it would take to extend that to an async iterator trait. As before, I am interested in exploring the “core capabilities” that would be needed to make everything work. Start somewhere: Just assume we want Box In the first post of this series, we talked about how invoking an async fn through a dyn trait should to have the return type of that async fn be a Box<dyn Future> — but only when calling it through a dyn type, not all the time.

In the previous “dyn async traits” posts, I talked about how we can think about the compiler as synthesizing an impl that performed the dynamic dispatch. In this post, I wanted to start explore a theoretical future in which this impl was written manually by the Rust programmer. This is in part a thought exercise, but it’s also a possible ingredient for a future design: if we could give programmers more control over the “impl Trait for dyn Trait” impl, then we could enable a lot of use cases.

In the previous post, we uncovered a key challenge for dyn and async traits: the fact that, in Rust today, dyn types have to specify the values for all associated types. This post is going to dive into more background about how dyn traits work today, and in particular it will talk about where that limitation comes from. Today: Dyn traits implement the trait In Rust today, assuming you have a “dyn-safe” trait DoTheThing , then the type dyn DoTheThing implements Trait.

Over the last few weeks, Tyler Mandry and I have been digging hard into what it will take to implement async fn in traits. Per the new lang team initiative process, we are collecting our design thoughts in an ever-evolving website, the async fundamentals initiative. If you’re interested in the area, you should definitely poke around; you may be interested to read about the MVP that we hope to stabilize first, or the (very much WIP) evaluation doc which covers some of the challenges we are still working out.