Axum Connect-Web

Brings the protobuf-based Connect-Web RPC framework to Rust via idiomatic Axum.

Axum Version

• axum-connect:0.3 works with axum:0.7
• axum-connect:0.2 works with axum:0.6

Features 🔍

• Integrates into existing Axum HTTP applications seamlessly
• Closely mirrors Axum's API
  • Extract State and other parts that impl RpcFromRequestParts just like with Axum.
  • Return any type that impl RpcIntoResponse just like Axum.
• Generated types and service handlers are strongly typed and...
• Handlers enforce semantically correct HTTP 'parts' access.
• Allows users to derive RpcIntoResponse and RpcFromRequestParts just like Axum.
  • Note: These must be derivatives of Axum's types because they are more restrictive; you're not dealing with arbitrary HTTP any more, you're speaking connect-web RPC over HTTP.
• Wrap connect-web error handling in idiomatic Axum/Rust.
• Codegen from *.proto files in a separate crate.
• All the other amazing benefits that come with Axum, like the community, documentation and performance!

Caution ⚠️

We use axum-connect in production, but I don't kow that anyone with more sense does. It's written in Rust which obviously offers some amazing compiler guarantees, but it's not well tested or battle-proven yet. Do what you will with that information.

Please let me know if you're using axum-connect! And open issues if you find a bug.

Getting Started 🤓

Prior knowledge with Protobuf (both the IDL and it's use in RPC frameworks) and Axum are assumed.

Dependencies 👀

You'll need 2 axum-connect crates, one for code-gen and one for runtime use. Because of how prost works, you'll also need to add it to your own project. You'll obviously also need axum and tokio.

# Note: axum-connect-build will fetch `protoc` for you.
cargo add --build axum-connect-build
cargo add axum-connect prost axum
cargo add tokio --features full

Protobuf File 🥱

Start by creating the obligatory 'hello world' proto service definition.

proto/hello.proto

syntax = "proto3";

package hello;

message HelloRequest { string name = 1; }

message HelloResponse { string message = 1; }

service HelloWorldService {
  rpc SayHello(HelloRequest) returns (HelloResponse) {}
}

Codegen 🤔

Use the axum_connect_codegen crate to generate Rust code from the proto IDL.

Currently all codegen is done by having the proto files locally on-disk, and using a build.rs file. Someday I hope to support more of Buf's idioms like Remote Code Gen.

build.rs

use axum_connect_build::{axum_connect_codegen, AxumConnectGenSettings};

fn main() {
    // This helper will use `proto` as the import path, and globs all .proto
    // files in the `proto` directory.
    //
    // Note that you might need to re-save the `build.rs` file after updating
    // a proto file to get rust-analyzer to pickup the change. I haven't put
    // time into looking for a fix to that yet.
    let settings = AxumConnectGenSettings::from_directory_recursive("proto")
        .expect("failed to glob proto files");

    axum_connect_codegen(settings).unwrap();
}

The Fun Part 😁

With the boring stuff out of the way, let's implement our service using Axum!

use async_stream::stream;
use axum::{extract::Host, Router};
use axum_connect::{futures::Stream, prelude::*};
use error::Error;
use proto::hello::*;
use tower_http::cors::CorsLayer;

// Take a peak at error.rs to see how errors work in axum-connect.
mod error;

mod proto {
    // Include the generated code in a `proto` module.
    //
    // Note: I'm not super happy with this pattern. I hope to add support to `protoc-gen-prost` in
    // the near-ish future instead see:
    // https://github.com/neoeinstein/protoc-gen-prost/issues/82#issuecomment-1877107220 That will
    // better align with Buf.build's philosophy. This is how it works for now though.
    pub mod hello {
        include!(concat!(env!("OUT_DIR"), "/hello.rs"));
    }
}

#[tokio::main]
async fn main() {
    // Build our application with a route. Note the `rpc` method which was added by `axum-connect`.
    // It expect a service method handler, wrapped in it's respective type. The handler (below) is
    // just a normal Rust function. Just like Axum, it also supports extractors!
    let app = Router::new()
        // A standard unary (POST based) Connect-Web request handler.
        .rpc(HelloWorldService::say_hello(say_hello_unary))
        // A GET version of the same thing, which has well-defined semantics for caching.
        .rpc(HelloWorldService::say_hello_unary_get(say_hello_unary))
        // A server-streaming request handler. Very useful when you need them!
        .rpc(HelloWorldService::say_hello_stream(stream_three_reponses));

    let listener = tokio::net::TcpListener::bind("127.0.0.1:3030")
        .await
        .unwrap();
    println!("listening on http://{:?}", listener.local_addr().unwrap());
    axum::serve(listener, app.layer(CorsLayer::very_permissive()))
        .await
        .unwrap();
}

/// The bread-and-butter of Connect-Web, a Unary request handler.
///
/// Just to demo error handling, I've chose to return a `Result` here. If your method is
/// infallible, you could just as easily return a `HellResponse` directly. The error type I'm using
/// is defined in `error.rs` and is worth taking a quick peak at.
///
/// Like Axum, both the request AND response types just need to implement RpcFromRequestParts` and
/// `RpcIntoResponse` respectively. This allows for a ton of flexibility in what your handlers
/// actually accept/return. This is a concept very core to Axum, so I won't go too deep into the
/// ideology here.
async fn say_hello_unary(Host(host): Host, request: HelloRequest) -> Result<HelloResponse, Error> {
    Ok(HelloResponse {
        message: format!(
            "Hello {}! You're addressing the hostname: {}.",
            request.name.unwrap_or_else(|| "unnamed".to_string()),
            host
        ),
    })
}

/// This is a server-streaming request handler. Much more rare to see one in the wild, but they
/// sure are useful when you need them! axum-connect has only partial support for everything
/// connect-web defines in server-streaming requests. For example, it doesn't define a way to
/// return trailers. I've never once actually needed them, so it feels weird to muddy the API just
/// to support such a niche use. Trailers are IMO the worst single decision gRPC made, locking them
/// into HTTP/2 forever. I'm not a fan -.-
///
/// You can however return a stream of anything that converts `RpcIntoResponse`, just like the
/// unary handlers. Again, very flexible. In this case I'm using the amazing `async-stream` crate
/// to make the code nice and readable.
async fn stream_three_reponses(
    Host(host): Host,
    request: HelloRequest,
) -> impl Stream<Item = HelloResponse> {
    stream! {
        yield HelloResponse { message: "Hello".to_string() };
        tokio::time::sleep(std::time::Duration::from_secs(1)).await;
        yield HelloResponse { message: request.name().to_string() };
        tokio::time::sleep(std::time::Duration::from_secs(1)).await;
        yield HelloResponse { message: format!("You're addressing the hostname: {}", host) };
        tokio::time::sleep(std::time::Duration::from_secs(1)).await;
    }
}

SEND IT 🚀

cargo run -p axum-connect-example

It's Connect RPCs, so you can use the Buf Studio to test things out! https://buf.build/studio/athilenius/axum-connect/main/hello.HelloWorldService/SayHello?target=http%3A%2F%2Flocalhost%3A3030

Request/Response Parts 🙍‍♂️

Both the request and response types are derived in axum-connect. This might seem redundant at first.

Let's go over the easier one first, RpcIntoResponse. Connect RPCs are not arbitrary HTML requests, they cannot return arbitrary HTML responses. For example, streaming responses MUST return an HTTP 200 regardless of error state. Keeping with Axum's (fantastic) paradigm, that is enforced by the type system. RPC handlers may not return arbitrary HTML, but instead must return something that axum-connect knows how to turn into a valid Connect response.

Somewhat less intuitively, axum-connect derives RpcFromRequestParts, which is almost identical to Axum's FromRequestParts. Importantly though, FromRequestParts can return back an error of arbitrary HTML responses, which is a problem for the same reason.

Axum also allows a FromRequest to occupy the last parameter in a handler which consumed the remainder of the HTTP request (including the body). axum-connect needs to handle the request input itself, so there is no equivalent for RPCs handlers.

Roadmap / Stated Non-Goals 🛣️

• Explore better typing than RpcFromRequestParts
  • Ideally clients only need to provide an RpcIntoError, but I haven't fully thought this problem through. I just know that having to specific both a FromRequestParts and an RpcFromRequestParts on custom types is a PITA.
• Rework error responses to allow for multiple errors
• Version checking between generated and runtime code
• A plan for forward-compatibility
• Bring everything in-line with connect-web and...
• Comprehensive integration tests

More Distant Goals 🌜

• I would love to also support a WASM-ready client library
• Use buf.build to support remote codegen and streamlined proto handling
• Support gRPC calls
  • I don't think this is hard to do, I just have no personal use-case for it
• Possibly maybe-someday support BiDi streaming over WebRTC
  • This would require connect-web picking up support for the same
  • WebRTC streams because they are DTLS/SRTP and are resilient
• Replace Prost (with something custom and simpler)

Non-goals 🙅

• To support every feature gRPC does
  • You get a lot of this already with Axum, but gRPC is a monster that I don't wish to reproduce. That complexity is useful for Google, and gets in the way for pretty much everyone else.
• To do everything and integrate with everything
  • I plan on keeping axum-connect highly focused. Good at what it does and nothing else.
  • This is idiomatic Rust. Do one thing well, and leave the rest to other crates.

Prost and Protobuf 📖

Protoc Version

The installed version of protoc can be configured in the AxumConnectGenSettings if you need/wish to do so. Setting the value to None will disable the download entirely.

Reasoning

Prost stopped shipping protoc binaries (a decision I disagree with) so axum-connect-build internally uses protoc-fetcher to download and resolve a copy of protoc. This is far more turnkey than forcing every build environment (often Heroku and/or Docker) to have a recent protoc binary pre-installed. This behavior can be disabled if you disagree, or if you need to comply with corporate policy, or your build environment is offline.

I would someday like to replace all of it with a new 'lean and mean' protoc library for the Rust community. One with a built-in parser, that supports only the latest proto3 syntax as well as the canonical JSON serialization format and explicitly doesn't support many of the rarely used features. But that day is not today.

License 🧾

Axum-Connect is dual licensed (at your option)

• MIT License (LICENSE-MIT or http://opensource.org/licenses/MIT)
• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)