# touché Touché is a low level but fully featured HTTP 1.0/1.1 library. It tries to mimic [hyper](https://crates.io/crates/hyper), but with a synchronous API. For now only the server API is implemented. ## Hello world ```rust no_run use touche::{Response, Server, StatusCode}; fn main() -> std::io::Result<()> { Server::bind("0.0.0.0:4444").serve(|_req| { Response::builder() .status(StatusCode::OK) .body("Hello World") }) } ``` ## Features - HTTP Server (thread per connection model, backed by a thread pool) - Non buffered (streaming) requests and response bodies - HTTP/1.1 pipelining - TLS - Upgrade connections - Trailers headers - 100 continue expectation - Unix sockets servers ## Comparison with Hyper Touché shares a lot of similarities with Hyper: - "Low level" - Uses the [http crate](https://crates.io/crates/http) to represent HTTP related types - Allows fine-grained implementations of streaming HTTP bodies But also has some key differences: - It is synchronous - Uses `Vec` to represent bytes instead of [Bytes](https://crates.io/crates/bytes) - Doesn't support HTTP 2 (and probably never will) ## Handling persistent connections with non blocking IO Connection-per-thread web servers are notorious bad with persistent connections like websockets or event streams. This is primarily because the thread gets locked to the connection until it is closed. One solution to this problem is to handle such connections with non-blocking IO. By doing so, the server thread becomes available for other connections. The following example demonstrates a single-threaded touché server that handles websockets upgrades to a Tokio runtime. ```rust no_run use std::{error::Error, sync::Arc}; use futures::{stream::StreamExt, SinkExt}; use tokio::{net::TcpStream, runtime}; use tokio_tungstenite::{tungstenite::protocol::Role, WebSocketStream}; use touche::{upgrade::Upgrade, Body, Connection, Request, Server}; fn main() -> std::io::Result<()> { let runtime = Arc::new(runtime::Builder::new_multi_thread().enable_all().build()?); Server::builder() .max_threads(1) .bind("0.0.0.0:4444") .serve(move |req: Request| { let runtime = runtime.clone(); let res = tungstenite::handshake::server::create_response(&req.map(|_| ()))?; Ok::<_, Box>(res.upgrade(move |stream: Connection| { let stream = stream.downcast::().unwrap(); stream.set_nonblocking(true).unwrap(); runtime.spawn(async move { let stream = TcpStream::from_std(stream).unwrap(); let mut ws = WebSocketStream::from_raw_socket(stream, Role::Server, None).await; while let Some(Ok(msg)) = ws.next().await { if msg.is_text() && ws.send(msg).await.is_err() { break; } } }); })) }) } ``` ## Other examples ### Chunked response ```rust no_run use std::{error::Error, thread}; use touche::{Body, Response, Server, StatusCode}; fn main() -> std::io::Result<()> { Server::bind("0.0.0.0:4444").serve(|_req| { let (channel, body) = Body::channel(); thread::spawn(move || { channel.send("chunk1").unwrap(); channel.send("chunk2").unwrap(); channel.send("chunk3").unwrap(); }); Response::builder() .status(StatusCode::OK) .body(body) }) } ``` ### Streaming files ```rust no_run use std::{fs, io}; use touche::{Body, Response, Server, StatusCode}; fn main() -> std::io::Result<()> { Server::bind("0.0.0.0:4444").serve(|_req| { let file = fs::File::open("./examples/file.rs")?; Ok::<_, io::Error>( Response::builder() .status(StatusCode::OK) .body(Body::try_from(file)?) .unwrap(), ) }) } ``` ### Naive routing with pattern matching ```rust no_run use touche::{body::HttpBody, Body, Method, Request, Response, Server, StatusCode}; fn main() -> std::io::Result<()> { Server::builder() .bind("0.0.0.0:4444") .serve(|req: Request| { match (req.method(), req.uri().path()) { (_, "/") => Response::builder() .status(StatusCode::OK) .body(Body::from("Usage: curl -d hello localhost:4444/echo\n")), // Responds with the same payload (&Method::POST, "/echo") => Response::builder() .status(StatusCode::OK) .body(req.into_body()), // Responds with the reversed payload (&Method::POST, "/reverse") => { let body = req.into_body().into_bytes().unwrap_or_default(); match std::str::from_utf8(&body) { Ok(message) => Response::builder() .status(StatusCode::OK) .body(message.chars().rev().collect::().into()), Err(err) => Response::builder() .status(StatusCode::BAD_REQUEST) .body(err.to_string().into()), } } _ => Response::builder() .status(StatusCode::NOT_FOUND) .body(Body::empty()), } }) } ``` ### Response upgrades ```rust no_run use std::io::{BufRead, BufReader, BufWriter, Write}; use touche::{header, upgrade::Upgrade, Body, Connection, Response, Server, StatusCode}; fn main() -> std::io::Result<()> { Server::bind("0.0.0.0:4444").serve(|_req| { Response::builder() .status(StatusCode::SWITCHING_PROTOCOLS) .header(header::UPGRADE, "line-protocol") .upgrade(|stream: Connection| { let reader = BufReader::new(stream.clone()); let mut writer = BufWriter::new(stream); // Just a simple protocol that will echo every line sent for line in reader.lines() { match line { Ok(line) if line.as_str() == "quit" => break, Ok(line) => { writer.write_all(format!("{line}\n").as_bytes()); writer.flush(); } Err(_err) => break, }; } }) .body(Body::empty()) }) } ``` You can find other examples in the [examples directory](https://github.com/reu/touche/tree/master/examples). ## Performance While the primary focus is having a simple and readable implementation, the library shows some decent performance. A simple benchmark of the hello_world.rs example gives the following result: ```sh $ cat /proc/cpuinfo | grep name | uniq model name : AMD Ryzen 5 5600G with Radeon Graphics $ wrk --latency -t6 -c 200 -d 10s http://localhost:4444 Running 10s test @ http://localhost:4444 6 threads and 200 connections Thread Stats Avg Stdev Max +/- Stdev Latency 153.37us 391.20us 19.41ms 99.37% Req/Sec 76.11k 13.21k 89.14k 82.67% Latency Distribution 50% 126.00us 75% 160.00us 90% 209.00us 99% 360.00us 4544074 requests in 10.01s, 225.35MB read Requests/sec: 454157.11 Transfer/sec: 22.52MB ``` The result is on par with Hyper's hello world running on the same machine. ## Disclaimer This library is by no means a critique to Hyper or to async Rust. I **really** love both of them. The main motivation I had to write this library was to be able to introduce Rust to my co-workers (which are mainly web developers). A synchronous library is way more beginner friendly than an async one, and by having an API that ressembles the "canonical" HTTP Rust library, people can learn Rust concepts in a easier way before adventuring through Hyper and async.