Add sgl_router_attempt_http_responses_total for single attempt information (#15037)

sgl-model-gateway/src/observability/metrics.rs

@@ -330,10 +330,11 @@ impl RouterMetrics {
     }
 
     // TODO unify metric names
-    pub fn record_upstream_http_response(route: &str, status_code: u16) {
+    pub fn record_attempt_http_response(route: &str, status_code: u16, error_code: &str) {
-        counter!("sgl_router_upstream_http_responses_total",
+        counter!("sgl_router_attempt_http_responses_total",
             "route" => route.to_string(),
-            "status_code" => status_code.to_string()
+            "status_code" => status_code.to_string(),
+            "error_code" => error_code.to_string()
         )
         .increment(1);
     }

sgl-model-gateway/src/routers/http/router.rs

@@ -170,8 +170,18 @@ impl Router {
             &self.retry_config,
             // operation per attempt
             |_: u32| async {
-                self.route_typed_request_once(headers, typed_req, route, model_id, is_stream, &text)
+                let res = self
-                    .await
+                    .route_typed_request_once(headers, typed_req, route, model_id, is_stream, &text)
+                    .await;
+
+                // Need to be outside `route_typed_request_once` because that function has multiple return paths
+                RouterMetrics::record_attempt_http_response(
+                    route,
+                    res.status().as_u16(),
+                    extract_error_code_from_response(&res),
+                );
+
+                res
             },
             // should_retry predicate
             |res, _attempt| is_retryable_status(res.status()),
@@ -505,8 +515,6 @@ impl Router {
             }
         };
 
-        RouterMetrics::record_upstream_http_response(route, res.status().as_u16());
-
         let status = StatusCode::from_u16(res.status().as_u16())
             .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
 
@@ -715,6 +723,8 @@ fn convert_reqwest_error(e: reqwest::Error) -> Response {
 
 use async_trait::async_trait;
 
+use crate::routers::error::extract_error_code_from_response;
+
 #[async_trait]
 impl RouterTrait for Router {
     fn as_any(&self) -> &dyn std::any::Any {