Why
Rust can help with writing high performance web servers with very consistent latencies. Which means that you can use smaller boxes to host the services, or can keep a monolithic application scaling for as long as possible, before jumping into microservices, which come with issues of their own.
Setup
cargo new warp-server
Use cargo add to add dependencies to the toml file. You'd have to install cargo edit, if you don't want to, you can edit, cargo.toml manually.
cargo add tokio warp serde
Cargo add doesn't set features, so you'll have to edit manually anyway. Serde will need the derive feature, and tokio the macros.
[package]
name = "warp-server"
version = "0.1.0"
edition = "2018"
[dependencies]
tokio = { version = "0.2.11", features = ["macros"] }
warp = "0.2.1"
serde = { version = "1.0.104", features = ["derive"] }
Then add a simple endpoint to the main.rs file. We'll do a simple rest health endpoint, that returns a hardcoded version.
use warp::Filter;
use warp::reply::json;
use serde::{ Serialize, Deserialize };
#[derive(Serialize, Deserialize)]
struct HealthResponse {
version: u32
}
#[tokio::main]
async fn main() {
let health = warp::path!("api" / "health")
.map(|| json(&HealthResponse { version: 42 }));
warp::serve(health)
.run(([127, 0, 0, 1], 3030))
.await;
}
Now if you run cargo run or cargo --release from the command line, and navigate to http://localhost:3030/api/health you should see
the health endpoint response.
> curl http://localhost:3030/api/health
{"version":42}
We can do a small bench with wrk2. We can see that the our small server has a very good latency. 1.6ms for 99 percentile is nothing to scoff at. You can look below for a simpler node server that does less work(no route matching and no json serialization) that has a higher latency. Obviously you should take these simple benchmarks with a grain of salt, I didn't disable my machine's turbo boost, didn't make sure no background jobs were running, this is not a good way to do benches.
> wrk -t2 -c100 -d20s -R10000 --latency http://127.0.0.1:3030/api/health
Running 20s test @ http://127.0.0.1:3030/api/health
2 threads and 100 connections
Thread calibration: mean lat.: 0.791ms, rate sampling interval: 10ms
Thread calibration: mean lat.: 0.786ms, rate sampling interval: 10ms
Thread Stats Avg Stdev Max +/- Stdev
Latency 743.63us 381.72us 2.23ms 61.29%
Req/Sec 5.13k 337.81 7.00k 77.67%
Latency Distribution (HdrHistogram - Recorded Latency)
50.000% 738.00us
75.000% 1.05ms
90.000% 1.24ms
99.000% 1.60ms
99.900% 2.06ms
99.990% 2.20ms
99.999% 2.23ms
100.000% 2.23ms
...
#[Mean = 0.744, StdDeviation = 0.382]
#[Max = 2.232, Total count = 97500]
#[Buckets = 27, SubBuckets = 2048]
----------------------------------------------------------
198797 requests in 20.00s, 23.13MB read
Requests/sec: 9939.65
Transfer/sec: 1.16MB
> cat index.js
const http = require('http');
http
.createServer(function(req, res) {
res.writeHead(200, {'Content-Type': 'application/json'});
res.end('{"version": 42}');
})
.listen(3030);
> wrk -t2 -c100 -d20s -R10000 --latency http://127.0.0.1:3030/api/health
Running 20s test @ http://127.0.0.1:3030/api/health
2 threads and 100 connections
Thread calibration: mean lat.: 1.178ms, rate sampling interval: 10ms
Thread calibration: mean lat.: 1.134ms, rate sampling interval: 10ms
Thread Stats Avg Stdev Max +/- Stdev
Latency 1.05ms 489.88us 3.16ms 63.85%
Req/Sec 5.29k 677.16 8.67k 82.52%
Latency Distribution (HdrHistogram - Recorded Latency)
50.000% 1.05ms
75.000% 1.41ms
90.000% 1.70ms
99.000% 2.14ms
99.900% 2.48ms
99.990% 3.03ms
99.999% 3.12ms
100.000% 3.16ms
#[Mean = 1.049, StdDeviation = 0.490]
#[Max = 3.160, Total count = 97500]
#[Buckets = 27, SubBuckets = 2048]
----------------------------------------------------------
194242 requests in 20.00s, 30.57MB read
Requests/sec: 9711.71
Transfer/sec: 1.53MB
RAM usage
Comparing the RAM usage(RES column in htop, which is memory usage in KB), we can see that the node server uses about ~43MB, while the warp server ~3.8MB, which is 11 times less. This means you can launch 10x more hello world services on single box with warp.
PID USER PRI NI VIRT RES SHR S CPU% MEM% TIME+ Command
19344 user 20 0 569M 43512 27128 S 0.0 0.3 0:05.76 node index.js
PID USER PRI NI VIRT RES SHR S CPU% MEM% TIME+ Command
19589 user 20 0 5456 3816 2356 S 0.0 0.0 0:02.01 target/release/warp-server
Conclusion
Warp is a very clean and fast web framework for Rust. It has support for http/1.1 http/2, tls, websockets and a whole lot more. If you want to write some servers that run blazing fast and consume next to no RAM, then warp is a good option.