fix(prometheus): serve /metrics directly so prometheus scrapers don't hit the 307 redirect (#30079)

app.mount("/metrics", ...) introduced in v1.88.0 (e23d06dda4) made
starlette emit a 307 'Location: /metrics/' for the bare /metrics path.
prometheus-compatible scrapers do not follow redirects, so they got an
empty body while the 'up' metric stayed 1 -- silent metric loss.

replace the mount with two explicit GET routes (/metrics and
/metrics/) that both call generate_latest(registry) directly. matches
prometheus_client.make_asgi_app's behaviour without the starlette
mount-redirect.

Fixes #30079

litellm/integrations/prometheus.py

@@ -3932,30 +3932,48 @@ class PrometheusLogger(CustomLogger):
     @staticmethod
     def _mount_metrics_endpoint():
         """
-        Mount the Prometheus metrics endpoint with optional authentication.
+        Expose the Prometheus metrics endpoint with optional authentication.
 
-        Args:
-            require_auth (bool, optional): Whether to require authentication for the metrics endpoint.
-                                        Defaults to False.
+        Registers explicit GET routes for ``/metrics`` and ``/metrics/``
+        instead of using ``app.mount(...)``. Starlette mounts emit a 307
+        redirect when the request path equals the mount prefix without a
+        trailing slash; Prometheus-compatible scrapers do not follow
+        redirects, so bare-path scrapes return an empty body and silently
+        record no metrics (#30079).
         """
-        from prometheus_client import make_asgi_app
+        from fastapi import Response
+        from prometheus_client import (
+            CONTENT_TYPE_LATEST,
+            generate_latest,
+        )
 
         from litellm._logging import verbose_proxy_logger
         from litellm.proxy.proxy_server import app
 
-        # Create metrics ASGI app
         if "PROMETHEUS_MULTIPROC_DIR" in os.environ:
             from prometheus_client import CollectorRegistry, multiprocess
 
             registry = CollectorRegistry()
             multiprocess.MultiProcessCollector(registry)
-            metrics_app = make_asgi_app(registry)
         else:
-            metrics_app = make_asgi_app()
+            from prometheus_client import REGISTRY as registry
 
-        # Mount the metrics app to the app
-        app.mount("/metrics", metrics_app)
-        verbose_proxy_logger.debug("Starting Prometheus Metrics on /metrics (no authentication)")
+        async def _metrics_endpoint() -> Response:
+            return Response(
+                content=generate_latest(registry),
+                media_type=CONTENT_TYPE_LATEST,
+            )
+
+        for path in ("/metrics", "/metrics/"):
+            app.add_api_route(
+                path,
+                _metrics_endpoint,
+                methods=["GET"],
+                include_in_schema=False,
+            )
+        verbose_proxy_logger.debug(
+            "Starting Prometheus Metrics on /metrics (no authentication)"
+        )
 
 
 def _prometheus_labels_from_context(

tests/test_litellm/integrations/test_prometheus_metrics_endpoint.py

@@ -0,0 +1,56 @@
+"""
+Regression for #30079.
+
+After 1.88.0 the Prometheus ``/metrics`` endpoint was mounted via
+``app.mount("/metrics", make_asgi_app())``. Starlette mounts emit a 307
+``Location: /metrics/`` when the request path equals the mount prefix
+without a trailing slash, and Prometheus / Grafana Alloy scrapers do not
+follow redirects -- so ``GET /metrics`` returns an empty body while the
+``up`` metric stays ``1``, hiding the regression.
+
+These tests assert that BOTH ``/metrics`` and ``/metrics/`` return 200
+directly with the Prometheus text-format content type, no 307 in between.
+"""
+
+import os
+
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+
+def test_metrics_endpoint_no_trailing_slash_no_redirect(monkeypatch):
+    monkeypatch.delenv("PROMETHEUS_MULTIPROC_DIR", raising=False)
+
+    fresh_app = FastAPI()
+    monkeypatch.setattr("litellm.proxy.proxy_server.app", fresh_app)
+
+    from litellm.integrations.prometheus import PrometheusLogger
+
+    PrometheusLogger._mount_metrics_endpoint()
+
+    client = TestClient(fresh_app, follow_redirects=False)
+
+    response = client.get("/metrics")
+
+    assert (
+        response.status_code == 200
+    ), f"bare /metrics must not 307 (got {response.status_code}); see #30079"
+    assert "text/plain" in response.headers.get("content-type", "")
+
+
+def test_metrics_endpoint_with_trailing_slash_works(monkeypatch):
+    monkeypatch.delenv("PROMETHEUS_MULTIPROC_DIR", raising=False)
+
+    fresh_app = FastAPI()
+    monkeypatch.setattr("litellm.proxy.proxy_server.app", fresh_app)
+
+    from litellm.integrations.prometheus import PrometheusLogger
+
+    PrometheusLogger._mount_metrics_endpoint()
+
+    client = TestClient(fresh_app, follow_redirects=False)
+
+    response = client.get("/metrics/")
+
+    assert response.status_code == 200
+    assert "text/plain" in response.headers.get("content-type", "")