fix(router): limit cost poll cooldown exemption to provider 404s

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/router.py

@@ -179,7 +179,7 @@ from litellm.router_utils.cooldown_handlers import (
     _get_cooldown_deployments,
     _set_cooldown_deployments,
     is_advisor_orchestration_failure,
-    is_background_response_cost_poll_failure,
+    is_background_response_cost_poll_not_found,
     is_caller_timeout_408,
 )
 from litellm.router_utils.fallback_event_handlers import (
@@ -8142,10 +8142,10 @@ class Router:
             litellm_params: Final = kwargs.get("litellm_params", {})
             _model_info: Final = litellm_params.get("model_info", {})
 
-            if is_background_response_cost_poll_failure(litellm_params):
+            if is_background_response_cost_poll_not_found(exception, litellm_params):
                 verbose_router_logger.debug(
                     "Router: Exiting 'deployment_callback_on_failure' without cooldown. "
-                    "Failure came from the background response cost poll, not the deployment's health."
+                    "Provider 404 came from the background response cost poll, not the deployment's health."
                 )
                 return False
 

litellm/router_utils/cooldown_handlers.py

@@ -64,12 +64,9 @@ def is_advisor_orchestration_failure(exception: BaseException | None) -> bool:
     return bool(getattr(exception, _ADVISOR_ORCHESTRATION_FAILURE_ATTR, False))
 
 
-def is_background_response_cost_poll_failure(litellm_params: Mapping[str, object]) -> bool:
-    """Whether the failed call was the enterprise cost poller reading back a stored background response.
-
-    A provider 404 there means the provider dropped the stored object, not that the deployment is unhealthy.
-    """
-    return any(
+def is_background_response_cost_poll_not_found(exception: Exception, litellm_params: Mapping[str, object]) -> bool:
+    """Whether a background response cost poll failed with a provider 404."""
+    return getattr(exception, "status_code", None) == 404 and any(
         isinstance(candidate, Mapping)
         and candidate.get(INTERNAL_CALL_ORIGIN_METADATA_KEY) == BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
         for candidate in (litellm_params.get("metadata"), litellm_params.get("litellm_metadata"))

tests/test_litellm/test_router.py

@@ -8789,6 +8789,7 @@ class TestBackgroundResponseCostPollCooldown:
                     "model_info": {"id": "dep-1"},
                 }
             ],
+            allowed_fails=0,
         )
 
     def _cooled_down_ids(self, router):
@@ -8801,16 +8802,13 @@ class TestBackgroundResponseCostPollCooldown:
         )
 
     def _deployment_callback_on_failure(self, router, kwargs):
-        import asyncio
         from datetime import datetime
 
-        async def callback():
-            now = datetime.now()
-            return router.deployment_callback_on_failure(kwargs, None, now, now)
+        now = datetime.now()
+        return router.deployment_callback_on_failure(kwargs, None, now, now)
 
-        return asyncio.run(callback())
-
-    def test_untagged_not_found_cools_down_deployment(self):
+    @pytest.mark.asyncio
+    async def test_untagged_not_found_cools_down_deployment(self):
         router = self._router()
         assert (
             self._deployment_callback_on_failure(
@@ -8856,7 +8854,33 @@ class TestBackgroundResponseCostPollCooldown:
         value = get_deployment_failures_for_current_minute(litellm_router_instance=router, deployment_id="dep-1")
         assert not value
 
-    def test_other_internal_origin_not_found_still_cools_down_deployment(self):
+    @pytest.mark.asyncio
+    async def test_cost_poll_non_404_still_cools_down_deployment(self):
+        from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
+        from litellm.types.utils import BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
+
+        router = self._router()
+        assert (
+            self._deployment_callback_on_failure(
+                router,
+                {
+                    "exception": litellm.InternalServerError(
+                        message="upstream 500", llm_provider="openai", model="gpt-4.1"
+                    ),
+                    "litellm_params": {
+                        "model_info": {"id": "dep-1"},
+                        "litellm_metadata": {
+                            INTERNAL_CALL_ORIGIN_METADATA_KEY: BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
+                        },
+                    },
+                },
+            )
+            is True
+        )
+        assert "dep-1" in self._cooled_down_ids(router)
+
+    @pytest.mark.asyncio
+    async def test_other_internal_origin_not_found_still_cools_down_deployment(self):
         from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
         from litellm.types.utils import AUTOROUTER_CLASSIFIER_CALL_ORIGIN