fix(router): skip cooldown for background response cost poll 404s

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

litellm/router.py

@@ -179,6 +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_caller_timeout_408,
 )
 from litellm.router_utils.fallback_event_handlers import (
@@ -8137,12 +8138,19 @@ class Router:
                 )
                 return False
 
-            exception_status: Final = getattr(exception, "status_code", "")
-
             # Cache litellm_params to avoid repeated dict lookups
             litellm_params: Final = kwargs.get("litellm_params", {})
             _model_info: Final = litellm_params.get("model_info", {})
 
+            if is_background_response_cost_poll_failure(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."
+                )
+                return False
+
+            exception_status: Final = getattr(exception, "status_code", "")
+
             if is_caller_timeout_408(kwargs, exception_status):
                 verbose_router_logger.debug(
                     "Router: Exiting 'deployment_callback_on_failure' without cooldown. "

litellm/router_utils/cooldown_handlers.py

@@ -20,9 +20,11 @@ from litellm.constants import (
     DEFAULT_COOLDOWN_TIME_SECONDS,
     DEFAULT_FAILURE_THRESHOLD_MINIMUM_REQUESTS,
     DEFAULT_FAILURE_THRESHOLD_PERCENT,
+    INTERNAL_CALL_ORIGIN_METADATA_KEY,
     SINGLE_DEPLOYMENT_TRAFFIC_FAILURE_THRESHOLD,
 )
 from litellm.router_utils.cooldown_callbacks import router_cooldown_event_callback
+from litellm.types.utils import BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
 
 from .router_callbacks.track_deployment_metrics import (
     get_deployment_failures_for_current_minute,
@@ -62,6 +64,18 @@ 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(
+        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"))
+    )
+
+
 _EXCEPTION_POLICY_FIELDS: Final[tuple[tuple[type, str], ...]] = (
     # ContentPolicyViolationError subclasses BadRequestError, so it must be checked first.
     (litellm.ContentPolicyViolationError, "ContentPolicyViolationErrorAllowedFails"),

tests/test_litellm/test_router.py

@@ -8779,6 +8779,104 @@ class TestAdvisorSubCallCooldown:
         assert "dep-1" not in self._cooled_down_ids(router)
 
 
+class TestBackgroundResponseCostPollCooldown:
+    def _router(self):
+        return litellm.Router(
+            model_list=[
+                {
+                    "model_name": "gpt-4.1",
+                    "litellm_params": {"model": "openai/gpt-4.1"},
+                    "model_info": {"id": "dep-1"},
+                }
+            ],
+        )
+
+    def _cooled_down_ids(self, router):
+        active = router.cooldown_cache.get_active_cooldowns(model_ids=["dep-1"], parent_otel_span=None)
+        return [entry[0] for entry in active]
+
+    def _not_found(self):
+        return litellm.NotFoundError(
+            message="Response with id 'resp_gone' not found.", llm_provider="openai", model="gpt-4.1"
+        )
+
+    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)
+
+        return asyncio.run(callback())
+
+    def test_untagged_not_found_cools_down_deployment(self):
+        router = self._router()
+        assert (
+            self._deployment_callback_on_failure(
+                router,
+                {
+                    "exception": self._not_found(),
+                    "litellm_params": {"model_info": {"id": "dep-1"}, "metadata": {}},
+                },
+            )
+            is True
+        )
+        assert "dep-1" in self._cooled_down_ids(router)
+
+    def test_cost_poll_not_found_does_not_cool_down_deployment(self):
+        from datetime import datetime
+
+        from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
+        from litellm.router_utils.router_callbacks.track_deployment_metrics import (
+            get_deployment_failures_for_current_minute,
+        )
+        from litellm.types.utils import BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
+
+        router = self._router()
+        now = datetime.now()
+        assert (
+            router.deployment_callback_on_failure(
+                {
+                    "exception": self._not_found(),
+                    "litellm_params": {
+                        "model_info": {"id": "dep-1"},
+                        "litellm_metadata": {
+                            INTERNAL_CALL_ORIGIN_METADATA_KEY: BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
+                        },
+                    },
+                },
+                None,
+                now,
+                now,
+            )
+            is False
+        )
+        assert self._cooled_down_ids(router) == []
+        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):
+        from litellm.constants import INTERNAL_CALL_ORIGIN_METADATA_KEY
+        from litellm.types.utils import AUTOROUTER_CLASSIFIER_CALL_ORIGIN
+
+        router = self._router()
+        assert (
+            self._deployment_callback_on_failure(
+                router,
+                {
+                    "exception": self._not_found(),
+                    "litellm_params": {
+                        "model_info": {"id": "dep-1"},
+                        "litellm_metadata": {INTERNAL_CALL_ORIGIN_METADATA_KEY: AUTOROUTER_CLASSIFIER_CALL_ORIGIN},
+                    },
+                },
+            )
+            is True
+        )
+        assert "dep-1" in self._cooled_down_ids(router)
+
+
 class TestCallerTimeoutCooldown:
     """A timeout the caller set (the proxy's `timeout` body field or x-litellm-timeout
     header) comes back as a 408 whatever the deployment's health, so it must neither