Lot 5 : /v1 relayé par la façade directement au moteur

La surface compatible OpenAI ne fait plus le détour par le Go : la façade
authentifie avec la clé des complétions (préflight CORS exempté), réécrit
le corps vers le second slot quand il est en service (id_slot à 1, valeurs
gardées telles quelles), retire les X-Forwarded-* et relaie au moteur sans
tampon — Host d'origine, sauf pour Strata.

Les décisions restent au Go, qui tient l'état de la conversation : la
façade lui demande où router, ouvre un ticket « en vol » avant de relayer
(refus pendant un benchmark ou une optimisation, avec Retry-After) et le
rend quand le flux se termine, que le client parte ou que le moteur soit
injoignable. Sans ticket obtenu, la requête est refusée : le Go pourrait
sinon effacer le slot en plein calcul.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
MichaelandClaude Opus 5.5 committed 2026-10-14 17:47:40 +02:00
1 parent 451770bfcf
commit c9c9ac1ddd
7 files changed
+646 -1

No files matched your search

+39
View File
@@ -24,6 +24,12 @@ pub struct FakeGo {
pub broken: Mutex<bool>,
/// Chemins relayés (hors API interne).
pub relayed: Mutex<Vec<String>>,
/// /v1 : port du faux moteur, second slot en service, refus à renvoyer
/// par begin, journal des begin/end.
pub engine_port: Mutex<u16>,
pub side_live: Mutex<bool>,
pub refuse: Mutex<Option<serde_json::Value>>,
pub oai_log: Mutex<Vec<String>>,
}
impl FakeGo {
@@ -95,6 +101,39 @@ async fn handle(st: Arc<FakeGo>, req: Request<Incoming>) -> Response<Full<Bytes>
let key = st.get("state", "web_key").unwrap_or_default();
resp(200, format!(r#"{{"web_key":"{key}","version":"0.15.0"}}"#))
}
"/internal/v1/oai/route" => {
let key = st.get("state", "api_key").unwrap_or_default();
let v = serde_json::json!({
"port": *st.engine_port.lock().unwrap(),
"strata": false,
"key": key,
"key_ok": true,
"max_body": 1024,
"side_live": *st.side_live.lock().unwrap(),
});
resp(200, v.to_string())
}
"/internal/v1/oai/begin" => {
let body = req.into_body().collect().await.unwrap().to_bytes();
let v: serde_json::Value = serde_json::from_slice(&body).unwrap();
st.oai_log.lock().unwrap().push(format!(
"begin read={} side={} post_v1={}",
v["read"], v["side"], v["post_v1"]
));
if let Some(r) = st.refuse.lock().unwrap().clone() {
return resp(200, serde_json::json!({ "refuse": r }).to_string());
}
resp(200, r#"{"ticket":7}"#)
}
"/internal/v1/oai/end" => {
let body = req.into_body().collect().await.unwrap().to_bytes();
let v: serde_json::Value = serde_json::from_slice(&body).unwrap();
st.oai_log
.lock()
.unwrap()
.push(format!("end {}", v["ticket"]));
resp(204, "")
}
"/internal/v1/store/all" => {
let s = st.store.lock().unwrap();
let m: BTreeMap<String, String> = s
+242
View File
@@ -0,0 +1,242 @@
//! Lot 5 : /v1 relayé par la façade directement au moteur, décisions au Go.
mod common;
use bytes::Bytes;
use common::fakego;
use http_body_util::{BodyExt, Full, StreamBody};
use hyper::body::{Frame, Incoming};
use hyper::{Request, Response, StatusCode, service::service_fn};
use hyper_util::client::legacy::Client;
use hyper_util::rt::{TokioExecutor, TokioIo};
use std::convert::Infallible;
use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};
use tokio::net::TcpListener;
type B = http_body_util::combinators::BoxBody<Bytes, Infallible>;
/// Faux moteur : /v1/models renvoie la requête vue (en-têtes + corps),
/// /v1/chat/completions diffuse deux événements à 500 ms d'écart.
async fn faux_moteur(vu: Arc<Mutex<Vec<String>>>) -> u16 {
let l = TcpListener::bind("127.0.0.1:0").await.unwrap();
let port = l.local_addr().unwrap().port();
tokio::spawn(async move {
while let Ok((s, _)) = l.accept().await {
let vu = vu.clone();
tokio::spawn(async move {
let svc = service_fn(move |req: Request<Incoming>| {
let vu = vu.clone();
async move {
let path = req.uri().path().to_string();
let mut lines: Vec<String> = req
.headers()
.iter()
.map(|(k, v)| format!("{k}: {}", v.to_str().unwrap()))
.collect();
let body = req.into_body().collect().await.unwrap().to_bytes();
lines.push(format!("body: {}", String::from_utf8_lossy(&body)));
vu.lock().unwrap().push(lines.join("\n"));
let r: Response<B> = if path == "/v1/chat/completions" {
let (tx, rx) = tokio::sync::mpsc::channel::<Bytes>(2);
tokio::spawn(async move {
tx.send(Bytes::from_static(b"data: a\n\n")).await.unwrap();
tokio::time::sleep(Duration::from_millis(500)).await;
let _ = tx.send(Bytes::from_static(b"data: b\n\n")).await;
});
let s = futures_util::stream::unfold(rx, |mut rx| async move {
rx.recv()
.await
.map(|b| (Ok::<_, Infallible>(Frame::data(b)), rx))
});
Response::new(BodyExt::boxed(StreamBody::new(s)))
} else {
Response::new(Full::new(Bytes::from_static(b"{\"data\":[]}")).boxed())
};
Ok::<_, Infallible>(r)
}
});
let _ = hyper::server::conn::http1::Builder::new()
.serve_connection(TokioIo::new(s), svc)
.await;
});
}
});
port
}
async fn send(
facade: std::net::SocketAddr,
method: &str,
path: &str,
headers: &[(&str, &str)],
body: &str,
) -> Response<Incoming> {
let mut req = Request::builder()
.method(method)
.uri(format!("http://{facade}{path}"));
for (k, v) in headers {
req = req.header(*k, *v);
}
let client: Client<_, Full<Bytes>> = Client::builder(TokioExecutor::new()).build_http();
client
.request(req.body(Full::new(Bytes::from(body.to_string()))).unwrap())
.await
.unwrap()
}
async fn attendre_end(go: &fakego::FakeGo) {
for _ in 0..50 {
if go
.oai_log
.lock()
.unwrap()
.iter()
.any(|l| l.starts_with("end"))
{
return;
}
tokio::time::sleep(Duration::from_millis(20)).await;
}
}
#[tokio::test]
async fn relais_et_ticket() {
let t = tempfile::tempdir().unwrap();
let (f, go, _stop) = fakego::start(t.path().into()).await;
let vu = Arc::new(Mutex::new(Vec::new()));
*go.engine_port.lock().unwrap() = faux_moteur(vu.clone()).await;
let r = send(
f,
"GET",
"/v1/models",
&[("host", "loki.exemple.fr"), ("x-forwarded-for", "6.6.6.6")],
"",
)
.await;
assert_eq!(r.status(), StatusCode::OK);
let body = r.into_body().collect().await.unwrap().to_bytes();
assert_eq!(&body[..], b"{\"data\":[]}");
attendre_end(&go).await;
assert_eq!(
*go.oai_log.lock().unwrap(),
vec!["begin read=true side=false post_v1=false", "end 7"]
);
let vu = vu.lock().unwrap()[0].clone();
assert!(vu.contains("host: loki.exemple.fr"), "{vu}");
assert!(!vu.contains("x-forwarded-for"), "{vu}");
assert!(go.relayed.lock().unwrap().is_empty(), "/v1 relayé au Go");
}
#[tokio::test]
async fn flux_sans_tampon_ticket_rendu_a_la_fin() {
let t = tempfile::tempdir().unwrap();
let (f, go, _stop) = fakego::start(t.path().into()).await;
*go.engine_port.lock().unwrap() = faux_moteur(Arc::new(Mutex::new(Vec::new()))).await;
let t0 = Instant::now();
let r = send(f, "POST", "/v1/chat/completions", &[], r#"{"stream":true}"#).await;
let mut body = r.into_body();
let first = body.frame().await.unwrap().unwrap().into_data().unwrap();
assert_eq!(&first[..], b"data: a\n\n");
assert!(
t0.elapsed() < Duration::from_millis(300),
"{:?}",
t0.elapsed()
);
// Flux en cours : pas encore de end.
assert!(
!go.oai_log
.lock()
.unwrap()
.iter()
.any(|l| l.starts_with("end"))
);
while body.frame().await.is_some() {}
attendre_end(&go).await;
assert_eq!(
*go.oai_log.lock().unwrap(),
vec!["begin read=false side=false post_v1=true", "end 7"]
);
}
#[tokio::test]
async fn second_slot_reecrit_le_corps() {
let t = tempfile::tempdir().unwrap();
let (f, go, _stop) = fakego::start(t.path().into()).await;
let vu = Arc::new(Mutex::new(Vec::new()));
*go.engine_port.lock().unwrap() = faux_moteur(vu.clone()).await;
*go.side_live.lock().unwrap() = true;
let r = send(f, "POST", "/v1/models", &[], r#"{"model":"m","id_slot":0}"#).await;
assert_eq!(r.status(), StatusCode::OK);
let vu = vu.lock().unwrap()[0].clone();
assert!(vu.contains(r#"body: {"id_slot":1,"model":"m"}"#), "{vu}");
assert!(go.oai_log.lock().unwrap()[0].contains("side=true"));
// Corps trop gros pour le second slot (max_body = 1024 dans le faux Go).
let gros = format!(r#"{{"x":"{}"}}"#, "a".repeat(2000));
let r = send(f, "POST", "/v1/models", &[], &gros).await;
assert_eq!(r.status(), StatusCode::PAYLOAD_TOO_LARGE);
}
#[tokio::test]
async fn cle_des_completions() {
let t = tempfile::tempdir().unwrap();
let (f, go, _stop) = fakego::start(t.path().into()).await;
*go.engine_port.lock().unwrap() = faux_moteur(Arc::new(Mutex::new(Vec::new()))).await;
go.set("state", "api_key", "cle-oai");
let r = send(f, "GET", "/v1/models", &[], "").await;
assert_eq!(r.status(), StatusCode::UNAUTHORIZED);
assert_eq!(r.headers()["www-authenticate"], r#"Bearer realm="loki""#);
let body = r.into_body().collect().await.unwrap().to_bytes();
let v: serde_json::Value = serde_json::from_slice(&body).unwrap();
assert_eq!(v["error"]["code"], "invalid_api_key");
assert!(
go.oai_log.lock().unwrap().is_empty(),
"begin sans authentification"
);
let r = send(
f,
"GET",
"/v1/models",
&[("authorization", "Bearer cle-oai")],
"",
)
.await;
assert_eq!(r.status(), StatusCode::OK);
// Préflight CORS : jamais d'Authorization, jamais refusé.
let r = send(f, "OPTIONS", "/v1/models", &[], "").await;
assert_eq!(r.status(), StatusCode::OK);
}
#[tokio::test]
async fn refus_benchmark() {
let t = tempfile::tempdir().unwrap();
let (f, go, _stop) = fakego::start(t.path().into()).await;
*go.engine_port.lock().unwrap() = faux_moteur(Arc::new(Mutex::new(Vec::new()))).await;
*go.refuse.lock().unwrap() = Some(serde_json::json!({
"status": 503, "message": "benchmark en cours sur ce moteur — réessaie dans un moment",
"type": "server_error", "code": "benchmark_running", "retry_after": "30"
}));
let r = send(f, "POST", "/v1/chat/completions", &[], "{}").await;
assert_eq!(r.status(), StatusCode::SERVICE_UNAVAILABLE);
assert_eq!(r.headers()["retry-after"], "30");
}
#[tokio::test]
async fn moteur_absent() {
let t = tempfile::tempdir().unwrap();
let (f, go, _stop) = fakego::start(t.path().into()).await;
let mort = TcpListener::bind("127.0.0.1:0")
.await
.unwrap()
.local_addr()
.unwrap()
.port();
*go.engine_port.lock().unwrap() = mort;
let r = send(f, "GET", "/v1/models", &[], "").await;
assert_eq!(r.status(), StatusCode::BAD_GATEWAY);
attendre_end(&go).await;
assert!(
go.oai_log.lock().unwrap().iter().any(|l| l == "end 7"),
"ticket non rendu"
);
}