[router] Count requests/responses at the HTTP edge for true intake (#29467)

experimental/sgl-router/monitoring/README.md

@@ -20,10 +20,11 @@ The dashboard graphs every family the router emits:
 
 | Metric | Type | What it shows |
 |---|---|---|
-| `sgl_router_requests_total` | Counter | Dispatches by `worker_url`, `model_id`, `mode`, `outcome` |
+| `sgl_router_requests_total` | Counter | **Edge intake** — every request received at the router HTTP boundary, by `route`, `method`, counted before worker dispatch (true intake) |
+| `sgl_router_responses_total` | Counter | **Edge responses** — every response returned, by `route`, `method`, `status_code` (incl. early-exit 400/413/503). `requests_total - responses_total` = received-but-not-answered |
+| `sgl_router_worker_requests_total` | Counter | Per-worker **dispatches** by `worker_url`, `model_id`, `mode`, `outcome` (recorded after dispatch; blind to pre-dispatch drops) |
 | `sgl_router_request_duration_seconds` | Histogram | End-to-end request latency by `model_id` |
 | `sgl_router_ttft_seconds` | Histogram | Time to first token (streaming) by `model_id` |
-| `sgl_router_responses_total` | Counter | Client-visible HTTP `status_code` |
 | `sgl_router_overlap_blocks` | Histogram | Cache-aware-zmq overlap blocks by `model_id` |
 | `sgl_router_active_load` | Gauge | Per-worker prefill-token / decode-block load |
 | `sgl_router_workers` | Gauge | Registered worker count by `mode` |

experimental/sgl-router/monitoring/grafana-dashboard.json

@@ -89,7 +89,7 @@
             "uid": "${datasource}"
           },
           "editorMode": "code",
-          "expr": "sum(rate(sgl_router_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
+          "expr": "sum(rate(sgl_router_worker_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
           "range": true,
           "refId": "A",
           "legendFormat": "__auto",
@@ -161,7 +161,7 @@
             "uid": "${datasource}"
           },
           "editorMode": "code",
-          "expr": "100 * sum(rate(sgl_router_requests_total{outcome=\"error\",model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval])) / clamp_min(sum(rate(sgl_router_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval])), 1e-9)",
+          "expr": "100 * sum(rate(sgl_router_worker_requests_total{outcome=\"error\",model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval])) / clamp_min(sum(rate(sgl_router_worker_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval])), 1e-9)",
           "range": true,
           "refId": "A",
           "legendFormat": "__auto",
@@ -407,7 +407,7 @@
             "uid": "${datasource}"
           },
           "editorMode": "code",
-          "expr": "sum by (outcome) (rate(sgl_router_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
+          "expr": "sum by (outcome) (rate(sgl_router_worker_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
           "range": true,
           "refId": "A",
           "legendFormat": "{{outcome}}",
@@ -500,7 +500,7 @@
             "uid": "${datasource}"
           },
           "editorMode": "code",
-          "expr": "sum by (mode) (rate(sgl_router_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
+          "expr": "sum by (mode) (rate(sgl_router_worker_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
           "range": true,
           "refId": "A",
           "legendFormat": "{{mode}}",
@@ -686,7 +686,7 @@
             "uid": "${datasource}"
           },
           "editorMode": "code",
-          "expr": "sum by (worker_url) (rate(sgl_router_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
+          "expr": "sum by (worker_url) (rate(sgl_router_worker_requests_total{model_id=~\"$model_id\",worker_url=~\"$worker_url\"}[$__rate_interval]))",
           "range": true,
           "refId": "A",
           "legendFormat": "{{worker_url}}",
@@ -2035,7 +2035,7 @@
           "type": "prometheus",
           "uid": "${datasource}"
         },
-        "definition": "label_values(sgl_router_requests_total, model_id)",
+        "definition": "label_values(sgl_router_worker_requests_total, model_id)",
         "hide": 0,
         "includeAll": true,
         "allValue": ".*",
@@ -2044,7 +2044,7 @@
         "name": "model_id",
         "options": [],
         "query": {
-          "query": "label_values(sgl_router_requests_total, model_id)",
+          "query": "label_values(sgl_router_worker_requests_total, model_id)",
           "refId": "StandardVariableQuery"
         },
         "refresh": 2,
@@ -2061,7 +2061,7 @@
           "type": "prometheus",
           "uid": "${datasource}"
         },
-        "definition": "label_values(sgl_router_requests_total, worker_url)",
+        "definition": "label_values(sgl_router_worker_requests_total, worker_url)",
         "hide": 0,
         "includeAll": true,
         "allValue": ".*",
@@ -2070,7 +2070,7 @@
         "name": "worker_url",
         "options": [],
         "query": {
-          "query": "label_values(sgl_router_requests_total, worker_url)",
+          "query": "label_values(sgl_router_worker_requests_total, worker_url)",
           "refId": "StandardVariableQuery"
         },
         "refresh": 2,

experimental/sgl-router/src/server/app.rs

@@ -3,7 +3,7 @@
 
 use crate::server::app_context::AppContext;
 use crate::server::routes::chat::MAX_CHAT_BODY_BYTES;
-use axum::extract::{DefaultBodyLimit, Request};
+use axum::extract::{DefaultBodyLimit, MatchedPath, Request, State};
 use axum::http::StatusCode;
 use axum::middleware::{self, Next};
 use axum::response::Response;
@@ -11,6 +11,25 @@ use axum::routing::{get, post};
 use axum::Router;
 use std::sync::Arc;
 
+/// Edge counters: `requests_total{route,method}` at entry (true intake, incl.
+/// requests parked/shed/cancelled before dispatch), `responses_total{...,
+/// status_code}` on exit (incl. early-exit 400/413/503). Their difference =
+/// received-but-not-answered, invisible to post-dispatch `worker_requests_total`.
+/// `route` is the matched template (not raw URI) to bound label cardinality.
+async fn count_requests(State(ctx): State<Arc<AppContext>>, req: Request, next: Next) -> Response {
+    let method = req.method().as_str().to_owned();
+    let route = req
+        .extensions()
+        .get::<MatchedPath>()
+        .map(|m| m.as_str().to_owned())
+        .unwrap_or_else(|| "unmatched".to_owned());
+    ctx.metrics.record_ingress(&route, &method);
+    let resp = next.run(req).await;
+    ctx.metrics
+        .record_response(&route, &method, resp.status().as_u16());
+    resp
+}
+
 /// Middleware: log 413 PAYLOAD_TOO_LARGE responses with the request method
 /// and URI so an operator investigating "client X gets 413s" has a
 /// server-side breadcrumb. The 413 is produced by axum's `DefaultBodyLimit`
@@ -57,5 +76,7 @@ pub fn build_router(ctx: Arc<AppContext>) -> Router {
             "/flush_cache",
             post(crate::server::routes::cache::flush_cache),
         )
+        // After routing, so MatchedPath is set for every route.
+        .layer(middleware::from_fn_with_state(ctx.clone(), count_requests))
         .with_state(ctx)
 }

experimental/sgl-router/src/server/metrics.rs

@@ -18,10 +18,11 @@
 //!
 //! | Metric | Type | Labels |
 //! |---|---|---|
-//! | `sgl_router_requests_total` | Counter | `worker_url`, `model_id`, `mode`, `outcome` |
+//! | `sgl_router_requests_total` | Counter | `route`, `method` |
+//! | `sgl_router_responses_total` | Counter | `route`, `method`, `status_code` |
+//! | `sgl_router_worker_requests_total` | Counter | `worker_url`, `model_id`, `mode`, `outcome` |
 //! | `sgl_router_request_duration_seconds` | Histogram | `model_id` |
 //! | `sgl_router_ttft_seconds` | Histogram | `model_id` |
-//! | `sgl_router_responses_total` | Counter | `status_code` |
 //! | `sgl_router_overlap_blocks` | Histogram | `model_id` |
 //! | `sgl_router_active_load` | Gauge | `worker_url`, `kind` |
 //! | `sgl_router_workers` | Gauge | `mode` |
@@ -203,14 +204,21 @@ impl ActiveLoadKind {
 /// internal state is `Arc`/`Atomic`/`Mutex`-protected.
 #[derive(Debug, Default)]
 pub struct MetricsRegistry {
-    requests_total: Mutex<HashMap<RequestKey, Arc<AtomicU64>>>,
+    // Edge counters (recorded at the app.rs middleware): intake at entry,
+    // responses at exit. `requests_total - responses_total` = received but
+    // never answered, which `worker_requests_total` (post-dispatch) can't see.
+    requests_total: Mutex<HashMap<EdgeKey, Arc<AtomicU64>>>,
+    responses_total: Mutex<HashMap<EdgeResponseKey, Arc<AtomicU64>>>,
+    // Per-worker dispatch outcomes (formerly `requests_total`). Recorded after
+    // dispatch, so blind to pre-dispatch drops; kept per-worker for the
+    // routing-convergence tests.
+    worker_requests_total: Mutex<HashMap<RequestKey, Arc<AtomicU64>>>,
     // Keyed by `model_id` only: a model's pool is either all-plain or all-PD
     // (the registry rejects mixed pools), so the worker `mode` would be a pure
     // function of `model_id` here — a redundant label. Per-worker `mode` lives
-    // on `requests_total` / the worker gauges instead.
+    // on `worker_requests_total` / the worker gauges instead.
     request_duration: Mutex<HashMap<String, Histogram>>,
     ttft_seconds: Mutex<HashMap<String, Histogram>>,
-    responses_total: Mutex<HashMap<u16, Arc<AtomicU64>>>,
     overlap_blocks: Mutex<HashMap<String, Histogram>>,
     active_load: Mutex<HashMap<ActiveLoadKey, Arc<AtomicI64>>>,
     stale_requests_total: Mutex<HashMap<&'static str, Arc<AtomicU64>>>,
@@ -227,6 +235,22 @@ struct RequestKey {
     outcome: &'static str,
 }
 
+/// Labels for the edge `requests_total` (intake) counter. `route` is the matched
+/// template (small fixed set), so cardinality is bounded.
+#[derive(Debug, Hash, Eq, PartialEq, Clone)]
+struct EdgeKey {
+    route: String,
+    method: String,
+}
+
+/// Labels for the edge `responses_total` counter: `EdgeKey` + final HTTP status.
+#[derive(Debug, Hash, Eq, PartialEq, Clone)]
+struct EdgeResponseKey {
+    route: String,
+    method: String,
+    status_code: u16,
+}
+
 /// Per-worker state sampled from the [`crate::workers::WorkerRegistry`] at
 /// scrape time and rendered as the `sgl_router_workers` /
 /// `sgl_router_worker_*` gauge families. Built by the `/metrics` route from
@@ -302,8 +326,25 @@ impl MetricsRegistry {
         Arc::new(Self::default())
     }
 
-    /// Bump `sgl_router_requests_total` for the given worker / model / mode / outcome.
+    /// Bump the edge intake counter `requests_total{route,method}`. Called at the
-    pub fn record_request(
+    /// middleware before worker pick, so it sees pre-dispatch drops.
+    pub fn record_ingress(&self, route: &str, method: &str) {
+        let key = EdgeKey {
+            route: route.to_owned(),
+            method: method.to_owned(),
+        };
+        let mut guard = self.requests_total.lock();
+        let counter = guard
+            .entry(key)
+            .or_insert_with(|| Arc::new(AtomicU64::new(0)))
+            .clone();
+        drop(guard);
+        counter.fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Bump `worker_requests_total`. Recorded after dispatch — see `record_ingress`
+    /// for true intake.
+    pub fn record_worker_request(
         &self,
         worker_url: &str,
         model_id: &str,
@@ -316,7 +357,7 @@ impl MetricsRegistry {
             mode: mode.as_str(),
             outcome: outcome.as_str(),
         };
-        let mut guard = self.requests_total.lock();
+        let mut guard = self.worker_requests_total.lock();
         let counter = guard
             .entry(key)
             .or_insert_with(|| Arc::new(AtomicU64::new(0)))
@@ -373,14 +414,18 @@ impl MetricsRegistry {
         hist.observe(seconds);
     }
 
-    /// Bump `sgl_router_responses_total{status_code}` for the HTTP status the
+    /// Bump the edge counter `responses_total{route,method,status_code}`. Called
-    /// client ultimately saw. Cardinality is bounded by the small set of
+    /// at the middleware, so it captures every outcome — incl. early-exit
-    /// status codes the router returns (2xx success, 4xx client, 5xx
+    /// 400/413/503 that the old per-handler site skipped.
-    /// upstream/proxy, 504 stale-cancel).
+    pub fn record_response(&self, route: &str, method: &str, status_code: u16) {
-    pub fn record_response(&self, status_code: u16) {
+        let key = EdgeResponseKey {
+            route: route.to_owned(),
+            method: method.to_owned(),
+            status_code,
+        };
         let mut guard = self.responses_total.lock();
         let counter = guard
-            .entry(status_code)
+            .entry(key)
             .or_insert_with(|| Arc::new(AtomicU64::new(0)))
             .clone();
         drop(guard);
@@ -472,12 +517,33 @@ impl MetricsRegistry {
     pub fn render_with_workers(&self, workers: &[WorkerSnapshot]) -> String {
         let mut out = String::new();
 
-        // requests_total
+        // requests_total — edge intake (every request, counted before dispatch)
         out.push_str(
-            "# HELP sgl_router_requests_total Total chat-completions requests dispatched to a worker.\n",
+            "# HELP sgl_router_requests_total Total requests received at the router HTTP edge, counted before worker dispatch (true intake).\n",
         );
         out.push_str("# TYPE sgl_router_requests_total counter\n");
         let guard = self.requests_total.lock();
+        let mut entries: Vec<(&EdgeKey, u64)> = guard
+            .iter()
+            .map(|(k, v)| (k, v.load(Ordering::Relaxed)))
+            .collect();
+        entries.sort_by(|a, b| (&a.0.route, &a.0.method).cmp(&(&b.0.route, &b.0.method)));
+        for (key, value) in entries {
+            out.push_str(&format!(
+                "sgl_router_requests_total{{route=\"{}\",method=\"{}\"}} {}\n",
+                escape_label(&key.route),
+                escape_label(&key.method),
+                value,
+            ));
+        }
+        drop(guard);
+
+        // worker_requests_total — per-worker dispatch outcomes (formerly requests_total)
+        out.push_str(
+            "# HELP sgl_router_worker_requests_total Chat-completions requests dispatched to a worker, by dispatch outcome.\n",
+        );
+        out.push_str("# TYPE sgl_router_worker_requests_total counter\n");
+        let guard = self.worker_requests_total.lock();
         // Sort for stable output — easier for tests.
         let mut entries: Vec<(&RequestKey, u64)> = guard
             .iter()
@@ -493,7 +559,7 @@ impl MetricsRegistry {
         });
         for (key, value) in entries {
             out.push_str(&format!(
-                "sgl_router_requests_total{{worker_url=\"{}\",model_id=\"{}\",mode=\"{}\",outcome=\"{}\"}} {}\n",
+                "sgl_router_worker_requests_total{{worker_url=\"{}\",model_id=\"{}\",mode=\"{}\",outcome=\"{}\"}} {}\n",
                 escape_label(&key.worker_url),
                 escape_label(&key.model_id),
                 key.mode,
@@ -538,21 +604,30 @@ impl MetricsRegistry {
         }
         drop(guard);
 
-        // responses_total
+        // responses_total — edge, by route/method/status (incl. early-exit 400/413/503)
         out.push_str(
-            "# HELP sgl_router_responses_total Chat-completions responses returned to clients, by HTTP status code (recorded after worker dispatch).\n",
+            "# HELP sgl_router_responses_total Responses returned at the router HTTP edge, by route, method and HTTP status code.\n",
         );
         out.push_str("# TYPE sgl_router_responses_total counter\n");
         let guard = self.responses_total.lock();
-        let mut entries: Vec<(u16, u64)> = guard
+        let mut entries: Vec<(&EdgeResponseKey, u64)> = guard
             .iter()
-            .map(|(k, v)| (*k, v.load(Ordering::Relaxed)))
+            .map(|(k, v)| (k, v.load(Ordering::Relaxed)))
             .collect();
-        entries.sort_by_key(|e| e.0);
+        entries.sort_by(|a, b| {
-        for (status_code, value) in entries {
+            (&a.0.route, &a.0.method, a.0.status_code).cmp(&(
+                &b.0.route,
+                &b.0.method,
+                b.0.status_code,
+            ))
+        });
+        for (key, value) in entries {
             out.push_str(&format!(
-                "sgl_router_responses_total{{status_code=\"{}\"}} {}\n",
+                "sgl_router_responses_total{{route=\"{}\",method=\"{}\",status_code=\"{}\"}} {}\n",
-                status_code, value,
+                escape_label(&key.route),
+                escape_label(&key.method),
+                key.status_code,
+                value,
             ));
         }
         drop(guard);
@@ -918,16 +993,35 @@ mod tests {
     }
 
     #[test]
-    fn record_response_counts_by_status_code() {
+    fn record_response_counts_by_route_method_status_code() {
         let reg = MetricsRegistry::new();
-        reg.record_response(200);
+        reg.record_response("/v1/chat/completions", "POST", 200);
-        reg.record_response(200);
+        reg.record_response("/v1/chat/completions", "POST", 200);
-        reg.record_response(502);
+        reg.record_response("/v1/chat/completions", "POST", 502);
-        reg.record_response(504);
+        reg.record_response("/v1/chat/completions", "POST", 504);
         let out = reg.render();
-        assert!(out.contains(r#"sgl_router_responses_total{status_code="200"} 2"#));
+        assert!(out.contains(
-        assert!(out.contains(r#"sgl_router_responses_total{status_code="502"} 1"#));
+            r#"sgl_router_responses_total{route="/v1/chat/completions",method="POST",status_code="200"} 2"#
-        assert!(out.contains(r#"sgl_router_responses_total{status_code="504"} 1"#));
+        ));
+        assert!(out.contains(
+            r#"sgl_router_responses_total{route="/v1/chat/completions",method="POST",status_code="502"} 1"#
+        ));
+        assert!(out.contains(
+            r#"sgl_router_responses_total{route="/v1/chat/completions",method="POST",status_code="504"} 1"#
+        ));
+    }
+
+    #[test]
+    fn record_ingress_counts_by_route_method() {
+        let reg = MetricsRegistry::new();
+        reg.record_ingress("/v1/chat/completions", "POST");
+        reg.record_ingress("/v1/chat/completions", "POST");
+        reg.record_ingress("/v1/models", "GET");
+        let out = reg.render();
+        assert!(out.contains(
+            r#"sgl_router_requests_total{route="/v1/chat/completions",method="POST"} 2"#
+        ));
+        assert!(out.contains(r#"sgl_router_requests_total{route="/v1/models",method="GET"} 1"#));
     }
 
     #[test]
@@ -981,15 +1075,15 @@ mod tests {
     }
 
     #[test]
-    fn record_request_emits_labelled_counter_line() {
+    fn record_worker_request_emits_labelled_counter_line() {
         let reg = MetricsRegistry::new();
-        reg.record_request(
+        reg.record_worker_request(
             "http://worker-a:30000",
             "tiny",
             WorkerModeLabel::Prefill,
             RequestOutcome::Success,
         );
-        reg.record_request(
+        reg.record_worker_request(
             "http://worker-a:30000",
             "tiny",
             WorkerModeLabel::Prefill,
@@ -997,7 +1091,7 @@ mod tests {
         );
         let out = reg.render();
         assert!(
-            out.contains(r#"sgl_router_requests_total{worker_url="http://worker-a:30000",model_id="tiny",mode="prefill",outcome="success"} 2"#),
+            out.contains(r#"sgl_router_worker_requests_total{worker_url="http://worker-a:30000",model_id="tiny",mode="prefill",outcome="success"} 2"#),
             "render did not include the expected counter line; got:\n{out}",
         );
     }
@@ -1111,7 +1205,7 @@ mod tests {
     #[test]
     fn label_values_escape_quotes_and_backslashes() {
         let reg = MetricsRegistry::new();
-        reg.record_request(
+        reg.record_worker_request(
             r#"http://"weird":30000"#,
             r"back\slash",
             WorkerModeLabel::Plain,

experimental/sgl-router/src/server/routes/chat.rs

@@ -544,7 +544,7 @@ pub async fn chat_completions(
         Err(_) => RequestOutcome::Error,
     };
     ctx.metrics
-        .record_request(&metrics_worker_url, &metrics_model, metrics_mode, outcome);
+        .record_worker_request(&metrics_worker_url, &metrics_model, metrics_mode, outcome);
 
     // Per-request access log — always on at INFO so incoming traffic and its
     // status are visible without DEBUG. `request_id` is the client/gateway
@@ -560,19 +560,19 @@ pub async fn chat_completions(
         Err(e) => e.status_code().as_u16(),
     };
 
-    // Record the client-visible HTTP status now that the outcome is known.
+    // Record end-to-end latency now that the outcome is known. Non-streaming:
-    // For non-streaming requests the body is already buffered here, so
+    // body is buffered here, so `start.elapsed()` is true e2e — record directly.
-    // `start.elapsed()` is true end-to-end latency — record it directly. For
+    // Streaming: body still pending, so the `RecordDurationOnDrop` guard in
-    // streaming, the body is still pending; the `RecordDurationOnDrop` guard
+    // `stream_guards` records it at stream completion (not just time-to-headers).
-    // packed into `stream_guards` records it at stream completion instead (so
+    // `elapsed` still feeds the access-log `latency_ms` for both.
-    // we don't capture only time-to-headers). `elapsed` still feeds the
+    //
-    // access-log `latency_ms` below for both.
+    // HTTP status is counted into `responses_total` by the edge middleware
+    // (app.rs), not here — the old per-handler site skipped early exits.
     let elapsed = start.elapsed();
     if !streaming {
         ctx.metrics
             .observe_request_duration(&metrics_model, elapsed.as_secs_f64());
     }
-    ctx.metrics.record_response(http_status);
     let outcome_str = match outcome {
         RequestOutcome::Success => "success",
         RequestOutcome::Error => "error",

experimental/sgl-router/src/server/routes/metrics.rs

@@ -102,7 +102,7 @@ mod tests {
     #[tokio::test]
     async fn metrics_endpoint_reflects_recorded_counters() {
         let ctx = Arc::new(AppContext::stub());
-        ctx.metrics.record_request(
+        ctx.metrics.record_worker_request(
             "http://w-test:30000",
             "tiny",
             WorkerModeLabel::Prefill,

experimental/sgl-router/tests/e2e/chat_completions/test_load_based_policy.py

@@ -19,7 +19,7 @@ _ACTIVE_RE = re.compile(
     r'^sgl_router_active_load\{worker_url="([^"]+)",kind="prefill_tokens"\}\s+(-?\d+)'
 )
 _REQ_TOTAL_RE = re.compile(
-    r"^sgl_router_requests_total\{([^}]*)\}\s+(\d+(?:\.\d+)?)\s*$"
+    r"^sgl_router_worker_requests_total\{([^}]*)\}\s+(\d+(?:\.\d+)?)\s*$"
 )
 _LABEL_RE = re.compile(r'(\w+)="([^"]*)"')
 

experimental/sgl-router/tests/e2e/chat_completions/test_two_router_convergence.py

@@ -74,7 +74,7 @@ PREFIX_Y = (_PREFIX_Y_BODY * 8).strip()
 
 
 _REQ_TOTAL_RE = re.compile(
-    r"^sgl_router_requests_total\{([^}]*)\}\s+(\d+(?:\.\d+)?)\s*$"
+    r"^sgl_router_worker_requests_total\{([^}]*)\}\s+(\d+(?:\.\d+)?)\s*$"
 )
 _LABEL_RE = re.compile(r'(\w+)="([^"]*)"')
 

experimental/sgl-router/tests/proxy/chat_routing.rs

@@ -90,6 +90,48 @@ async fn non_streaming_returns_200() {
     assert_eq!(v["choices"][0]["message"]["content"], "ok");
 }
 
+/// Edge counters fire through the real middleware: `requests_total` at entry +
+/// `responses_total` on exit, with matched-route/method labels. The unit tests
+/// call record_* directly, so this is the only check that the middleware is
+/// actually wired (MatchedPath -> record_ingress / record_response).
+#[tokio::test]
+async fn edge_counters_recorded_through_middleware() {
+    let worker = crate::common::mock_worker::MockWorker::start(vec![]).await;
+    let ctx = build_ctx_with_worker(&worker.url);
+    let app = build_router(ctx.clone());
+
+    let req = Request::builder()
+        .method("POST")
+        .uri("/v1/chat/completions")
+        .header("content-type", "application/json")
+        .body(Body::from(
+            serde_json::to_vec(&serde_json::json!({
+                "model": "tiny",
+                "messages": [{"role": "user", "content": "hi"}],
+                "stream": false
+            }))
+            .unwrap(),
+        ))
+        .unwrap();
+    let res = app.oneshot(req).await.unwrap();
+    assert_eq!(res.status(), StatusCode::OK);
+    let _ = res.into_body().collect().await;
+
+    let m = ctx.metrics.render();
+    // intake — counted at entry by the middleware (the path unit tests miss)
+    assert!(
+        m.contains(r#"sgl_router_requests_total{route="/v1/chat/completions",method="POST"} 1"#),
+        "edge intake counter missing; got:\n{m}",
+    );
+    // response — counted on the way out by the middleware
+    assert!(
+        m.contains(
+            r#"sgl_router_responses_total{route="/v1/chat/completions",method="POST",status_code="200"} 1"#
+        ),
+        "edge response counter missing; got:\n{m}",
+    );
+}
+
 #[tokio::test]
 async fn non_streaming_upstream_unreachable_returns_502_unreachable() {
     // Bind a port, drop it — guarantees a closed/refused TCP destination.
@@ -288,7 +330,9 @@ async fn streaming_5xx_request_records_duration_and_status_but_not_ttft() {
         "TTFT must NOT be recorded for a non-2xx streaming response; got:\n{m}",
     );
     assert!(
-        m.contains(r#"sgl_router_responses_total{status_code="500"} 1"#),
+        m.contains(
+            r#"sgl_router_responses_total{route="/v1/chat/completions",method="POST",status_code="500"} 1"#
+        ),
         "the 500 status must be counted; got:\n{m}",
     );
     assert!(

experimental/sgl-router/tests/proxy/sticky_routing.rs

@@ -93,12 +93,12 @@ fn chat_request(header: Option<(&str, &str)>) -> Request<Body> {
         .unwrap()
 }
 
-/// Parse `sgl_router_requests_total{...,outcome="success"} N` lines into a
+/// Parse `sgl_router_worker_requests_total{...,outcome="success"} N` lines into a
 /// map of worker_url -> success count.
 fn success_counts(metrics: &str) -> std::collections::HashMap<String, u64> {
     let mut counts = std::collections::HashMap::new();
     for line in metrics.lines() {
-        let Some(rest) = line.strip_prefix("sgl_router_requests_total{") else {
+        let Some(rest) = line.strip_prefix("sgl_router_worker_requests_total{") else {
             continue;
         };
         if !rest.contains(r#"outcome="success""#) {