[Bugfix][Router] Return 503 when failover ends on a backend client error

src/tests/test_instance_failover.py

@@ -1,6 +1,7 @@
 import json
 from unittest.mock import MagicMock, patch
 
+import aiohttp
 import pytest
 
 from vllm_router.routers.routing_logic import (
@@ -157,6 +158,26 @@ async def fail(*a, **kw):
             await route_general_request(req, "/v1/chat/completions", MagicMock())
 
 
+@pytest.mark.asyncio
+async def test_unreachable_backends_return_503_after_all_attempts(setup):
+    req, router = setup
+
+    async def fail(*a, **kw):
+        raise aiohttp.ClientConnectionError("down")
+        yield
+
+    with patch(
+        "vllm_router.services.request_service.request.process_request", side_effect=fail
+    ):
+        from vllm_router.services.request_service.request import route_general_request
+
+        response = await route_general_request(req, "/v1/chat/completions", MagicMock())
+
+    assert response.status_code == 503
+    assert json.loads(response.body) == {"error": "Failed to connect to backend: down"}
+    assert "X-Request-Id" in response.headers
+
+
 @pytest.mark.asyncio
 async def test_no_retry_when_disabled(setup):
     req, router = setup

src/vllm_router/services/request_service/request.py

@@ -699,6 +699,15 @@ async def route_general_request(
             )
 
     if last_error:
+        if isinstance(last_error, aiohttp.ClientError):
+            # Same status and body as proxy_multipart_request for an unreachable backend.
+            if tracing_active:
+                end_span(span, error=last_error, status_code=503)
+            return JSONResponse(
+                status_code=503,
+                content={"error": f"Failed to connect to backend: {last_error}"},
+                headers={"X-Request-Id": request_id},
+            )
         end_span(span, error=last_error, status_code=500) if tracing_active else None
         raise last_error