Async Interview #2: cramertj
For the second async interview, I spoke with Taylor Cramer – or
cramertj, as I’ll refer to him. cramertj is a member of the compiler
and lang teams and was – until recently – working on Fuchsia at
Google. He’s been a key player in Rust’s Async I/O design and in the
discussions around it. He was also responsible for a lot of the
implementation work to make
async fn a reality.
You can watch the video on YouTube. I’ve also embedded a copy here for your convenience:
Spreading this out over a few posts
So, cramertj and I had a long conversation, with a lot of technical detail. I was trying to get this blog post finished by last Friday but it took a lot of time! I decided it’s probably too much material to post in one go, so I’m going to break up the blog post into a few pieces (I’ll post the whole video though).
The blog post is mostly covering what cramertj had to say, though in some cases I’m also adding in various bits of background information or my own editorialization. I’m trying to mark it when I do that. =)
We kicked off the discussion talking a bit about the particulars of the Fuchsia project. Fuchsia is a microkernel architecture and thus a lot of the services one finds in a typical kernel are implemented as independent Fuchsia processes. These processes are implemented in Rust and use Async I/O.
Fuchsia uses its own unique executor and runtime
Because Fuchsia is not a unix system, its kernel primitives, like sockets and events, work quite differently. Fuchsia therefore uses its own custom executor and runtime, rather than building on a separate stack like tokio or async-std.
Fuchsia benefits from interoperability
Even though Fuchsia uses its own executor, it is able to reuse a lot of libraries from the ecosystem. For example, Fuchsia uses Hyper for its HTTP parsing. This is possible because Hyper offers a generic interface based on traits that Fuchsia can implement.
In general, cramertj feels that the best way to achieve interop is to offer trait-based interfaces. There are other projects, for example, that offer feature flags (e.g., to enable “tokio” compatibilty etc), but this tends to be a suboptimal way of managing things, at least for libraries.
For one thing, offer features means that support for systems like fuschia must be “upstreamed” into the project, whereas offering traits means that downsteam systems can implement the traits themselves.
In addition, using features to choose between alternatives can cause problems across larger dependency graphs. Features are always meant to be “additive” – i.e,. you can add any number of them – but features that choose between backends tend to be exclusive – i.e., you must choose at most one. This is a problem because cargo likes to take the union of all features across a dependency graph, and so having exclusive features can lead to miscompilations when things are combined.
Background topic: futures crate
cramertj and I next talked some about the futures crate. Before going much further into that, I want to give a bit of background on the futures crate itself and how its setup.
The futures crate has been very carefully setup to permit its components to evolve with minimal breakage and incompatibility across the ecosystem. However, my experience from talking to people has been that there is a lot of confusion as to how the futures crate is setup and why, and just how much they can rely on things not to change. So I want to spend a bit of time documenting my understanding the setup and its motivations.
futures crate has served as a kind of
experimental “proving ground” for various aspects of the future
design, including the
Future trait itself (which is now in std).
Currently, the futures crate is at version 0.3, and it offers a number of different categories of functionality:
- key traits like
- key primitives like [“async-aware” locks]
- traditional locks
- “extension” traits like
AsyncReadExt, and so forth
- these traits offer convenient combinator methods like
mapthat are not part of the corresponding base traits
- these traits offer convenient combinator methods like
- useful macros like
- useful bits of code such as a
ThreadPoolfor “off-loading” heavy computations
In fact, the first item in that list (“key traits”) is quite distinct
from the remaining items. In particular, if you are writing a library,
those key traits are things that you might well like to have in your
public interface. For example, if you are writing a parser that
operates on a stream of data, it might take a
AsyncRead as its
data source (just as a synchronous parser would take a
The remaining items on the list fall generally into the category of
“implementation details”. They ought to be “private” dependencies of
your crate. For example, you may use methods from
internally, but you don’t require other crates to use them; similarly
join! futures internally, but that is not something that
would show up in a function signature.
the futures crate is really a facade
One thing you’ll notice if you look more closely at the
crate is that it is in fact composed of a number of smaller crates.
futures crate itself simply ‘re-exports’ items from these
futures-core– defines the
Streamtrait (also the
Futuretrait, but that is an alias for std)
futures-io– defines the
futures-util– defines extension traits like
The goal of this facade is to permit things to evolve without forcing
semver-incompatible changes. For example, if the
should evolve, we might be forced to issue a new major version of
futures-io and thus ultimately issue a new
(say, 0.4). However, the version number of
unchanged. This means that if your crate only depends on the
Stream trait, it will be interoperable across both
and 0.4, since both of those versions are in fact re-exporting the
Stream trait (from
futures-core, whose version has not
In fact, if you are a library crate, it probably behooves you to avoid
depending on the
futures crate at all, and instead to declare
finer-grained dependencies; this will make it very clear whe you need
to declare a new semver release yourself.
cramertj: the best place for “standard” traits is in std
So, background aside, let me return to my discussion with
cramertj. One of the points that cramertj is that the only “truly
standard” place for a trait to live is libstd. Therefore, cramertj
feels like the next logical step for traits like
AsyncRead is to start moving them into the standard library. Once
they are there, this would be the strongest possible signal that
people can rely on them not to change.
we can move to libstd without breakage
You may be wondering what it would mean if we moved one of the traits
futures crate into libstd – would things in the
ecosystem that are currently using
futures have to update? The
answer is no, not necessarily.
Presuming that some trait from
futures is moved wholesale into
libstd (i.e., without any modification), then it is possible for us
to simply issue a new minor version of the
futures crate (and
the appropriate subcrate). This new minor version would change from
defining a trait (say,
Stream) to re-exporting the version from std.
As a concrete example, if we moved
to libstd (as cramertj advocates for later on), then we would issue a
0.3.2 release of
futures-io. This release would replace
AsyncRead with a
pub use that re-exports
AsyncRead from std. Now,
any crate in the ecosystem that previously depended on
0.3.1 can be
transparently upgraded to
0.3.2 (it’s a semver-compatibly change,
after all)1, and suddenly all references to
AsyncRead would be referencing the version from std. (This is, in
fact, exactly what happened with the futures trait; in 0.3.1., it is
simply re-exported from libcore.)
on the extension traits
One of the interesting points that cramertj made, though not until
later in the interview, is that when it comes to futures there are a
number of “smaller design decisions” one might make when it comes to
combinators. For example, consider a function like
As defined in the future crates, this function returns a “future to a
boolean”, so it has a signature like:
impl FnMut(&Item) -> impl Future<Output = bool>
This is effectively an async closure; I’ll summarize what cramertj had to say about async closures in one of the upcoming blog posts. However, you might plausibly wish instead to have a signature that just returns a boolean directly, like so:
impl FnMut(&Item) -> bool
For this reason, cramertj felt that it may make sense not to add these sorts of utilities into the standard library (or at least not yet), and instead to leave those extension traits in “user space”. Maybe when we have more experience we’ll be able to say what the best definition would be for the standard library.
(If I may editorialize, I do think it’s important that we add these sorts of helper methods to std eventually; even if there’s no single best choice, we should make some decisions, because it’ll be quite annoying to force everything to pull in utility crates for simple things.)
OK, that wraps it up for the first post. I have two more coming. In
the next post, we’ll discuss the design of the
AsyncWrite traits, and what we might want to change there. In
the final post, we’ll discuss async closures.
There is a thread on the Rust users forum for this series.
This change relies on the fact that cargo will generally not compile two distinct minor versions of a crate; so all crates that depend on
0.3.1would be compiled against