To match local conventions I needed to adjust the metric names and I'm not able to modify the scrape config. Thus, I made this patch which allows these AXUM_HTTP_* constants to be changed via envs.

It is common for applications metrics to have an application-specific prefix. Metrics can be renamed at scrape time of course, but in some environments an application is expected to merely expose metrics on a given port and no other configuration is allowed.

This commit allows the metric names to be changed at compile time by setting the following environment variables:

• AXUM_HTTP_REQUESTS_TOTAL
• AXUM_HTTP_REQUESTS_DURATION_SECONDS
• AXUM_HTTP_REQUESTS_PENDING

In the absence of these environment variables, the metric names are unchanged.

This commit also documents the usage of these envs and suggests setting them in .cargo/config.toml

Trying to address #4.

Added options to skip reporting route patterns to Prometheus, and also an option to group routes into one endpoint. Added a builder struct to be able to easily customize these. For more info, see the docs in the PR.

This is a draft, because I'm not sure whether this is a viable solution.

Hi,

I noticed that the metric axum_http_requests_pending does not get properly decremented when the client connection is aborted, so its value may increase over time.

I can reproduce with this test:

use axum::{routing::get, Router};
use axum_prometheus::PrometheusMetricLayer;
use tokio::{
    select,
    time::{sleep, Duration},
};

#[tokio::test]
async fn metrics_socket() {
    let (metrics, prom) = PrometheusMetricLayer::pair();
    let app: Router = Router::new()
        .route(
            "/test",
            get(|| async { sleep(Duration::from_millis(10)).await }),
        )
        .layer(metrics);

    assert_eq!(prom.render(), "");

    let (tx, rx) = tokio::sync::oneshot::channel();
    let server = tokio::spawn(async {
        axum::Server::bind(&([127, 0, 0, 1], 3000).into())
            .serve(app.into_make_service())
            .with_graceful_shutdown(async {
                rx.await.unwrap();
            })
            .await
            .unwrap();
    });

    select! {
        r = reqwest::get("http://localhost:3000/test") => {r.unwrap();},
        _ = sleep(Duration::from_millis(10)) => (),
    }

    sleep(Duration::from_millis(10)).await;
    assert_eq!(
        prom.render()
            .lines()
            .filter(|l| l.starts_with("axum_http_requests_pending"))
            .map(|l| l.split(" ").last().unwrap())
            .collect::<String>(),
        "0"
    );
    tx.send(()).unwrap();
    server.await.unwrap();
}

or by stress-testing with autocannon and observing the metric value is not 0 after it ends

(also the axum_http_requests_duration_seconds histogram is not updated, but i'm unsure which should be the correct behaviour for aborted requests here)

Hi, is it possible to label the metrics by configured route instead of endpoint?

e.g. if I have a route with path parameters (e.g. /foo/:bar/:baz) I want to get one set of metrics (or optionally one per allowed value of :bar and :baz by adding additional labels) instead of one per actual endpoint hit by anyone

currently there is a possible problem with exploding metrics if path params have high cardinality, either intentionally or someone trying to break our APIs