Tiny add sglang:http_requests_active metric (#16479)

python/sglang/srt/utils/common.py

@@ -1502,7 +1502,7 @@ def add_prometheus_middleware(app):
 
 
 def add_prometheus_track_response_middleware(app):
-    from prometheus_client import Counter
+    from prometheus_client import Counter, Gauge
 
     http_request_counter = Counter(
         name="sglang:http_requests_total",
@@ -1516,6 +1516,13 @@ def add_prometheus_track_response_middleware(app):
         labelnames=["endpoint", "status_code", "method"],
     )
 
+    http_requests_active = Gauge(
+        name="sglang:http_requests_active",
+        documentation="Number of currently active HTTP requests",
+        labelnames=["endpoint", "method"],
+        multiprocess_mode="livesum",
+    )
+
     @app.middleware("http")
     async def track_http_status_code(request, call_next):
         # With recording all requests, we have the risk of high cardinality if requests have arbitrary unhandled paths.
@@ -1524,7 +1531,9 @@ def add_prometheus_track_response_middleware(app):
         method = request.method
 
         http_request_counter.labels(endpoint=path, method=method).inc()
+        http_requests_active.labels(endpoint=path, method=method).inc()
 
+        try:
             response = await call_next(request)
 
             http_response_counter.labels(
@@ -1534,6 +1543,8 @@ def add_prometheus_track_response_middleware(app):
             ).inc()
 
             return response
+        finally:
+            http_requests_active.labels(endpoint=path, method=method).dec()
 
 
 # https://github.com/blueswen/fastapi-observability/blob/132a3c576f8b09e5311c68bd553215013bc75685/fastapi_app/utils.py#L98

test/srt/test_metrics.py

@@ -127,6 +127,7 @@ class TestEnableMetrics(CustomTestCase):
             "sglang:time_to_first_token_seconds",
             "sglang:inter_token_latency_seconds",
             "sglang:e2e_request_latency_seconds",
+            "sglang:http_requests_active",
         ]
         for metric in essential_metrics:
             self.assertIn(metric, metrics_text, f"Missing metric: {metric}")