diff --git a/litellm/proxy/common_utils/openai_error_payload.py b/litellm/proxy/common_utils/openai_error_payload.py index 3c91e87ee51..5b413626b3b 100644 --- a/litellm/proxy/common_utils/openai_error_payload.py +++ b/litellm/proxy/common_utils/openai_error_payload.py @@ -89,9 +89,9 @@ def proxy_exception_for(exc: Exception, default_status_code: int) -> ProxyExcept code=status.HTTP_503_SERVICE_UNAVAILABLE, ) case HTTPException(status_code=status_code, detail=detail): - return _openai_shaped_proxy_exception(exc, str(detail), status_code) + return openai_shaped_proxy_exception(exc, str(detail), status_code) case _: - return _openai_shaped_proxy_exception( + return openai_shaped_proxy_exception( exc, str(exc) or type(exc).__name__, error_status_code(exc, default_status_code) ) @@ -131,7 +131,7 @@ def _is_failover_refusal(exc: Exception) -> bool: return any(refusal in message for refusal in _FAILOVER_REFUSALS) -def _openai_shaped_proxy_exception(exc: Exception, message: str, status_code: int) -> ProxyException: +def openai_shaped_proxy_exception(exc: Exception, message: str, status_code: int) -> ProxyException: answered_status: Final = int(proxy_exception_status_code(message, str(status_code))) return ProxyException( message=message, diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py index 9f58c267231..0bbb49eb8a3 100644 --- a/litellm/proxy/utils.py +++ b/litellm/proxy/utils.py @@ -85,8 +85,8 @@ from litellm.proxy._types import ( from litellm.proxy.bug_report_config import build_proxy_bug_report from litellm.proxy.common_utils.openai_error_payload import ( attribute_of, - litellm_call_id_headers, - openai_error_param, + error_status_code, + openai_shaped_proxy_exception, with_litellm_call_id, ) from litellm.proxy.spend_tracking.spend_log_error_logger import spend_log_error @@ -9179,39 +9179,30 @@ def _recreate_writer_on_read_only_transaction(prisma_client: "PrismaClient | Non asyncio.create_task(prisma_client.recreate_read_only_writer(reason="postgres_read_only_transaction")) +def _client_message(e: Exception) -> str: + status_is_none: Final = hasattr(e, "status_code") and attribute_of(e, "status_code") is None + return "Internal server error" if status_is_none else strip_bug_report_notice(str(e)) + + def handle_exception_on_proxy(e: Exception, litellm_call_id: str | None = None) -> ProxyException: """ Returns an Exception as ProxyException, this ensures all exceptions are OpenAI API compatible """ - from fastapi import status - verbose_proxy_logger.exception("Exception: %s", e) if PrismaDBExceptionHandler.is_read_only_transaction_error(e): from litellm.proxy.proxy_server import prisma_client _recreate_writer_on_read_only_transaction(prisma_client) - headers: Final = litellm_call_id_headers(litellm_call_id) - if isinstance(e, HTTPException): - return ProxyException( - message=getattr(e, "detail", f"error({e})"), - type=ProxyErrorTypes.internal_server_error, - param=openai_error_param(e), - headers=headers, - code=getattr(e, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR), - ) - elif isinstance(e, ProxyException): + if isinstance(e, ProxyException): return with_litellm_call_id(e, litellm_call_id) - _status_code: Final = getattr(e, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR) + if isinstance(e, HTTPException): + return with_litellm_call_id(openai_shaped_proxy_exception(e, str(e.detail), e.status_code), litellm_call_id) if should_report_bug(e): verbose_proxy_logger.error(bug_report_notice(build_proxy_bug_report(e))) - return ProxyException( - message=strip_bug_report_notice(str(e)), - type=ProxyErrorTypes.internal_server_error, - param=openai_error_param(e), - headers=headers, - code=_status_code, - ) + message: Final = _client_message(e) + status_code: Final = error_status_code(e, status.HTTP_500_INTERNAL_SERVER_ERROR) + return with_litellm_call_id(openai_shaped_proxy_exception(e, message, status_code), litellm_call_id) def require_enterprise_license(feature: str | None = None) -> None: diff --git a/tests/integration/authorization/test_guardrail_submissions.py b/tests/integration/authorization/test_guardrail_submissions.py index f120028f46d..07d9a96c979 100644 --- a/tests/integration/authorization/test_guardrail_submissions.py +++ b/tests/integration/authorization/test_guardrail_submissions.py @@ -432,7 +432,7 @@ def test_team_guardrail_submissions_require_admin_approval( f"Guardrail '{first_name}' not found. " "Please ensure the guardrail is configured in your LiteLLM proxy." ), - "type": "internal_server_error", + "type": "invalid_request_error", "param": None, "code": "404", } @@ -481,7 +481,7 @@ def test_team_guardrail_submissions_require_admin_approval( f"Guardrail '{second_name}' not found. " "Please ensure the guardrail is configured in your LiteLLM proxy." ), - "type": "internal_server_error", + "type": "invalid_request_error", "param": None, "code": "404", } diff --git a/tests/integration/authorization/test_user_self_update_guard.py b/tests/integration/authorization/test_user_self_update_guard.py index d058b3c6998..a4ba3d8fb6c 100644 --- a/tests/integration/authorization/test_user_self_update_guard.py +++ b/tests/integration/authorization/test_user_self_update_guard.py @@ -279,7 +279,7 @@ def test_new_user_with_admin_role_requires_proxy_admin(gateway: Gateway, role: s assert refused.json() == { "error": { "message": f"Only proxy admins can create administrative users (proxy_admin, proxy_admin_viewer). Attempted to create user with role: {role}. Your role: internal_user", - "type": "internal_server_error", + "type": "auth_error", "param": None, "code": "403", } diff --git a/tests/integration/management/test_user_delete_revokes_keys.py b/tests/integration/management/test_user_delete_revokes_keys.py index 324963a0da9..edbbf550871 100644 --- a/tests/integration/management/test_user_delete_revokes_keys.py +++ b/tests/integration/management/test_user_delete_revokes_keys.py @@ -170,7 +170,7 @@ def test_user_delete_revokes_keys_and_memberships(gateway: Gateway, peer: Gatewa assert info.json() == { "error": { "message": f"User {user_id} not found", - "type": "internal_server_error", + "type": "invalid_request_error", "param": None, "code": "404", } @@ -337,7 +337,7 @@ def test_scim_user_delete_revokes_keys(gateway: Gateway, peer: Gateway, monkeypa expected_info: Final = { "error": { "message": f"User {user_id} not found", - "type": "internal_server_error", + "type": "invalid_request_error", "param": None, "code": "404", } diff --git a/tests/integration/observability/test_apply_guardrail_endpoint.py b/tests/integration/observability/test_apply_guardrail_endpoint.py index 9fa7d1a74db..f50679bd54e 100644 --- a/tests/integration/observability/test_apply_guardrail_endpoint.py +++ b/tests/integration/observability/test_apply_guardrail_endpoint.py @@ -156,7 +156,7 @@ def test_apply_guardrail_returns_a_block_error_for_generic_guardrails(gateway: G assert error.model_dump() == { "error": { "message": _BLOCK_REASON, - "type": "internal_server_error", + "type": "invalid_request_error", "param": None, "code": 400, } diff --git a/tests/unit/proxy/credential_endpoints/test_endpoints.py b/tests/unit/proxy/credential_endpoints/test_endpoints.py index 9dd36a535e4..2eeb9bf265d 100644 --- a/tests/unit/proxy/credential_endpoints/test_endpoints.py +++ b/tests/unit/proxy/credential_endpoints/test_endpoints.py @@ -3275,3 +3275,20 @@ class TestCredentialDisplayName: assert (by_name["display_name"], by_name["source"]) == ("Prod OpenAI", "db") assert config_by_name["source"] == "config" assert listed["openai-prod"]["credential_values"]["api_key"] != "sk-old" + + +def test_looking_up_a_missing_credential_is_labelled_a_client_error_not_a_gateway_failure( + restore_credential_list: None, +): + """Naming a credential that does not exist is the caller's own mistake, and the shared + handler used to label every one of these internal_server_error, so an SDK branching on + the type read a 404 it can never retry past as a gateway outage and backed off forever.""" + response: Final = _call_as("GET", "/credentials/by_name/definitely-not-there") + + assert response.status_code == 404, f"lookup answered {response.status_code}: {response.text}" + assert response.json()["error"] == { + "message": "Credential not found. Got credential name: definitely-not-there", + "type": "invalid_request_error", + "param": None, + "code": "404", + } diff --git a/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py b/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py index 859824e4a72..2aa56e7949b 100644 --- a/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py +++ b/tests/unit/proxy/management_endpoints/test_key_management_endpoints.py @@ -18736,7 +18736,7 @@ async def test_list_keys_rejects_invalid_expires(): expires="expred", ) - assert exc_info.value.code == "400" + assert (exc_info.value.code, exc_info.value.type) == ("400", "invalid_request_error") assert "Invalid expires value" in str(exc_info.value.message) diff --git a/tests/unit/proxy/utils/helpers/test_error_helpers.py b/tests/unit/proxy/utils/helpers/test_error_helpers.py index df399c8b1d2..b20857dbdaf 100644 --- a/tests/unit/proxy/utils/helpers/test_error_helpers.py +++ b/tests/unit/proxy/utils/helpers/test_error_helpers.py @@ -1,13 +1,20 @@ import asyncio import json +from typing import Final from unittest.mock import AsyncMock, MagicMock +import httpx +import openai import pytest from fastapi import HTTPException from prisma.errors import DataError +import litellm from litellm.proxy._types import ProxyErrorTypes, ProxyException +from litellm.proxy.common_request_processing import proxy_exception_from_route_error +from litellm.proxy.common_utils.openai_error_payload import proxy_exception_for from litellm.proxy.utils import get_error_message_str, handle_exception_on_proxy +from litellm.types.router import RouterErrors def normalize(value): @@ -96,7 +103,7 @@ def test_handle_exception_on_proxy_happy_path_http_exception(): assert snapshot == { "is_proxy_exception": True, "message": "forbidden", - "type": ProxyErrorTypes.internal_server_error.value, + "type": "auth_error", "code": "403", } @@ -156,7 +163,7 @@ def test_handle_exception_on_proxy_uses_attached_status_code_when_present(): assert snapshot == { "code": "418", "message": "teapot", - "type": ProxyErrorTypes.internal_server_error.value, + "type": "invalid_request_error", } @@ -176,6 +183,176 @@ def test_handle_exception_on_proxy_error_path_none_input_wraps_as_500(): } +@pytest.mark.parametrize( + "status_code, expected_type", + [ + (400, "invalid_request_error"), + (401, "auth_error"), + (403, "auth_error"), + (404, "invalid_request_error"), + (408, "invalid_request_error"), + (409, "invalid_request_error"), + (422, "invalid_request_error"), + (429, "rate_limit_error"), + (500, ProxyErrorTypes.internal_server_error.value), + (503, ProxyErrorTypes.internal_server_error.value), + ], +) +def test_handle_exception_on_proxy_labels_an_http_exception_by_the_status_it_answers( + status_code: int, expected_type: str +): + """A management route that rejects a caller's own argument answers 400, and calling that + internal_server_error tells the caller's retry loop the gateway is down when nothing but + their request is wrong.""" + result: Final = handle_exception_on_proxy(HTTPException(status_code=status_code, detail="boom")) + + assert (result.code, result.type) == (str(status_code), expected_type) + + +def _exception_carrying_status(carried: int | None, message: str = "boom") -> Exception: + class _CarriedStatus(Exception): + status_code = carried + + return _CarriedStatus(message) + + +@pytest.mark.parametrize( + "status_code, expected_type", + [ + (400, "invalid_request_error"), + (404, "invalid_request_error"), + (408, "invalid_request_error"), + (429, "rate_limit_error"), + (500, ProxyErrorTypes.internal_server_error.value), + ], +) +def test_handle_exception_on_proxy_labels_a_carried_status_the_same_way(status_code: int, expected_type: str): + """Routes also reach this handler with plain exceptions carrying their own status, and + those must be named by that status rather than all landing on the 500 label.""" + result: Final = handle_exception_on_proxy(_exception_carrying_status(status_code)) + + assert (result.code, result.type) == (str(status_code), expected_type) + + +def test_handle_exception_on_proxy_keeps_the_type_the_exception_already_names(): + """A rate limit litellm raised already names itself, and overwriting that with the status + map's guess would drop what the client branches on.""" + result: Final = handle_exception_on_proxy( + litellm.RateLimitError(message="slow down", llm_provider="openai", model="gpt-4o") + ) + + assert (result.code, result.type) == ("429", "throttling_error") + + +@pytest.mark.parametrize( + "exc", + [ + openai.InternalServerError( + message="upstream down", + response=httpx.Response(500, request=httpx.Request("GET", "https://api.openai.com/v1/batches/b")), + body={"type": "server_error"}, + ), + litellm.Timeout(message="took too long", model="gpt-6.1-sol", llm_provider="openai"), + litellm.RateLimitError(message="slow down", llm_provider="openai", model="gpt-6.1-sol"), + litellm.AuthenticationError(message="bad key", llm_provider="openai", model="gpt-6.1-sol"), + litellm.BadRequestError(message="bad input", model="gpt-6.1-sol", llm_provider="openai"), + HTTPException(status_code=404, detail="no such batch"), + ], + ids=["provider_5xx_naming_its_own_type", "timeout", "rate_limit", "bad_key", "bad_request", "http_404"], +) +def test_handle_exception_on_proxy_labels_a_failure_the_way_the_chat_route_does(exc: Exception): + """Batch, file, and management routes answer through this handler while chat answers through the route + error path, and one client calling both must read the same type for the same failure.""" + shared: Final = handle_exception_on_proxy(exc) + chat: Final = proxy_exception_from_route_error(exc) + + assert (shared.code, shared.type) == (chat.code, chat.type) + + +@pytest.mark.parametrize( + "exc", + [ + HTTPException(status_code=400, detail="bad budget_duration"), + HTTPException(status_code=401, detail="not allowed"), + HTTPException(status_code=403, detail="admin only"), + HTTPException(status_code=404, detail="no such team"), + litellm.AuthenticationError(message="bad key", llm_provider="openai", model="gpt-6.1-sol"), + ValueError(f"{RouterErrors.no_deployments_with_tag_routing.value}=['paid']"), + _exception_carrying_status(409, "duplicate credential"), + ], + ids=["http_400", "http_401", "http_403", "http_404", "provider_401", "tag_denial", "carried_409"], +) +def test_handle_exception_on_proxy_labels_a_failure_the_way_route_catch_alls_do(exc: Exception): + """Management routes answer some failures through their own catch-all and the rest through this handler, + so a client reading /key/list and /key/generate must see the same type and status for the same failure.""" + shared: Final = handle_exception_on_proxy(exc) + catch_all: Final = proxy_exception_for(exc, default_status_code=500) + + assert (shared.code, shared.type, shared.message) == (catch_all.code, catch_all.type, catch_all.message) + + +@pytest.mark.parametrize( + "message, expected", + [ + (f"{RouterErrors.no_deployments_with_provider_budget_routing.value}, try later", ("429", "rate_limit_error")), + (f"{RouterErrors.no_deployments_with_tag_routing.value}=['paid']", ("401", "auth_error")), + ], + ids=["no_deployment_left", "no_deployment_for_tags"], +) +def test_handle_exception_on_proxy_names_a_router_failure_by_the_status_it_answers( + message: str, expected: tuple[str, str] +): + """A router failure that carries no status answers the 429 or 401 its message maps to, and a 429 labelled + internal_server_error reads as an outage to a client that branches on the type.""" + result: Final = handle_exception_on_proxy(ValueError(message)) + + assert (result.code, result.type) == expected + + +def test_handle_exception_on_proxy_keeps_the_path_a_rejected_request_names(): + """A provider rejecting a batch input names the caller's own bucket path, and redacting the /tmp part of it + leaves the caller nothing to fix their request with.""" + result: Final = handle_exception_on_proxy( + litellm.BadRequestError( + message="Invalid input_file_id gs://my-bucket/tmp/batch_input.jsonl", + model="gemini-3.1-pro-preview", + llm_provider="vertex_ai", + ) + ) + + assert result.code == "400" + assert "gs://my-bucket/tmp/batch_input.jsonl" in result.message + + +def test_handle_exception_on_proxy_keeps_the_message_an_exception_with_a_status_carries(): + """Main sends a status-bearing failure's message verbatim, and relabeling its type must not change that text.""" + result: Final = handle_exception_on_proxy(_exception_carrying_status(503, "Cannot connect to host 10.12.4.7:8443")) + + assert (result.code, result.message) == ("503", "Cannot connect to host 10.12.4.7:8443") + + +@pytest.mark.parametrize( + "message", + [ + "boom", + "Cannot connect to host 10.12.4.7:8443", + "Cannot connect to host guardrail.internal.svc.cluster.local:443", + "Cannot connect to host [fd00::12]:8443", + ], + ids=["plain", "ipv4", "cluster_dns", "ipv6"], +) +def test_handle_exception_on_proxy_answers_a_generic_500_for_an_exception_whose_status_is_none(message: str): + """A guardrail whose upstream refuses the connection raises with status_code=None, which crashed the handler + into a 503 blaming the auth database, and its text names private hosts no virtual key was ever sent.""" + result: Final = handle_exception_on_proxy(_exception_carrying_status(None, message)) + + assert (result.code, result.type, result.message) == ( + "500", + ProxyErrorTypes.internal_server_error.value, + "Internal server error", + ) + + @pytest.mark.parametrize( "exc", [