Merge pull request #41059 from BerriAI/litellm_prometheus_pre_call_rate_limit_api_provider

fix(prometheus): label pre-call rate limit failures with the resolved api_provider

litellm/integrations/prometheus.py

@@ -41,6 +41,7 @@ from litellm.proxy._types import (
     LiteLLM_UserTable,
     UserAPIKeyAuth,
 )
+from litellm.proxy.hooks.rate_limiter_utils import PROXY_LLM_PROVIDER_FALLBACK
 from litellm.repositories.base_repository import BaseRepository
 from litellm.repositories.organization_repository import OrganizationRepository
 from litellm.repositories.team_repository import TeamRepository
@@ -2581,6 +2582,15 @@ class PrometheusLogger(CustomLogger):
             )
             return None
 
+    @staticmethod
+    def _extract_api_provider_from_exception(exception: Exception) -> str | None:
+        if not isinstance(exception, litellm.exceptions.RateLimitError):
+            return None
+        llm_provider: Final = exception.llm_provider
+        if not llm_provider or llm_provider == PROXY_LLM_PROVIDER_FALLBACK:
+            return None
+        return llm_provider
+
     async def async_post_call_failure_hook(
         self,
         request_data: dict,
@@ -2616,7 +2626,9 @@ class PrometheusLogger(CustomLogger):
             _metadata: Final = request_data.get("metadata", {}) or {}
             model_id: Final = _metadata.get("model_info", {}).get("id") or request_data.get("model_info", {}).get("id")
             rate_limit_category, rate_limit_type = self._extract_rate_limit_labels(original_exception)
-            api_provider: Final = self._extract_api_provider_from_request_data(request_data)
+            api_provider: Final = self._extract_api_provider_from_request_data(
+                request_data
+            ) or self._extract_api_provider_from_exception(original_exception)
             enum_values: Final = UserAPIKeyLabelValues(
                 end_user=user_api_key_dict.end_user_id,
                 user=user_api_key_dict.user_id,

tests/test_litellm/integrations/test_prometheus_labels.py

@@ -716,6 +716,77 @@ async def test_failure_hook_emits_api_provider_value_on_failed_requests_metric()
         _clear_prometheus_registry()
 
 
+async def _failed_requests_api_provider_labels(
+    request_data: dict[str, object],
+    original_exception: Exception,
+) -> list[str]:
+    from litellm.integrations.prometheus import PrometheusLogger
+    from litellm.proxy._types import UserAPIKeyAuth
+
+    _clear_prometheus_registry()
+    try:
+        await PrometheusLogger().async_post_call_failure_hook(
+            request_data=request_data,
+            original_exception=original_exception,
+            user_api_key_dict=UserAPIKeyAuth(token="tok"),
+        )
+        return [
+            s.labels.get("api_provider")
+            for s in _collected_samples("litellm_proxy_failed_requests_metric_total")
+        ]
+    finally:
+        _clear_prometheus_registry()
+
+
+@pytest.mark.asyncio
+async def test_failure_hook_emits_api_provider_from_pre_call_rate_limit_error_for_router_alias():
+    """
+    Pre-call limiters reject before a deployment lands on request_data and a
+    router alias cannot be inferred from its name, so the provider the limiter
+    resolved onto the exception is the only source for the label.
+    """
+    from litellm.exceptions import RateLimitType
+    from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError
+
+    err = ProxyRateLimitError(
+        detail={"error": "rpm exceeded"},
+        rate_limit_type=RateLimitType.REQUESTS,
+        model="openai/gpt-5.4-mini",
+        llm_provider="openai",
+    )
+
+    assert await _failed_requests_api_provider_labels(
+        {"model": "team-chat-model", "metadata": {}}, err
+    ) == ["openai"]
+
+
+@pytest.mark.asyncio
+async def test_failure_hook_leaves_api_provider_unset_when_rate_limiter_could_not_resolve_provider():
+    from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError
+
+    err = ProxyRateLimitError(detail={"error": "rpm exceeded"}, model="unknown-alias")
+
+    assert await _failed_requests_api_provider_labels(
+        {"model": "unknown-alias", "metadata": {}}, err
+    ) == ["None"]
+
+
+@pytest.mark.asyncio
+async def test_failure_hook_prefers_request_data_provider_over_exception_provider():
+    from litellm.exceptions import RateLimitError
+
+    err = RateLimitError(message="upstream 429", llm_provider="openai", model="gpt-4o")
+
+    assert await _failed_requests_api_provider_labels(
+        {
+            "model": "gpt-4o",
+            "metadata": {},
+            "litellm_params": {"custom_llm_provider": "azure"},
+        },
+        err,
+    ) == ["azure"]
+
+
 if __name__ == "__main__":
     test_user_email_in_required_metrics()
     test_user_email_label_exists()