fix(mcp): map a DB outage during bridge key reload to a retryable 503

get_key_object's raw transport error propagated uncaught out of
_reload_admitted_key as an opaque 500; classify it via the shared
_raise_503_if_db_unavailable helper (also used by the live-policy gate) so a
database outage is a retryable 503, while a key-not-found ProxyException stays
the fail-closed 401.
This commit is contained in:
Tin Chi Lo 2026-07-11 12:46:09 -07:00
parent ea64ef7a2a
commit f5f03cbd63
2 changed files with 40 additions and 7 deletions

View file

@ -581,11 +581,28 @@ class MCPRequestHandler:
)
except (ProxyException, HTTPException):
raise HTTPException(status_code=401, detail="Invalid or expired credential") from None
except Exception as e: # noqa: BLE001 # a DB outage during reload is a retryable 503, not an opaque 500
MCPRequestHandler._raise_503_if_db_unavailable(e)
raise
if not MCPRequestHandler._admitted_key_is_active(key_object):
raise HTTPException(status_code=401, detail="Invalid or expired credential")
await MCPRequestHandler._reject_if_admitted_owner_scim_deactivated(key_object)
return key_object
@staticmethod
def _raise_503_if_db_unavailable(e: Exception) -> None:
"""Raise a retryable 503 when ``e`` means the auth database is unreachable, else return so the
caller applies its own fail-closed mapping. A DB outage must not masquerade as an auth failure
(401) or surface as an opaque 500; the caller retries. Mirrors ``UserAPIKeyAuthExceptionHandler``,
which renders a service-unavailable database error as 503 on the standard pipeline."""
from litellm.proxy.db.exception_handler import PrismaDBExceptionHandler
if PrismaDBExceptionHandler.is_database_service_unavailable_error(e):
raise HTTPException(
status_code=503,
detail="Service Unavailable, the authentication database is temporarily unreachable. Please retry shortly.",
) from None
@staticmethod
async def _reject_if_admitted_owner_scim_deactivated(key_object: UserAPIKeyAuth) -> None:
"""Fail closed with a 401 when the key's owning user was deactivated via SCIM.
@ -636,8 +653,6 @@ class MCPRequestHandler:
it told an over-budget but validly-authenticated caller their credential was invalid,
which on a DCR client reads as broken auth and can trigger a pointless re-authorize
loop that cannot fix a budget problem, and it masked a DB outage as an auth error."""
from litellm.proxy.db.exception_handler import PrismaDBExceptionHandler
try:
await _run_centralized_common_checks(
user_api_key_auth_obj=admitted,
@ -650,11 +665,7 @@ class MCPRequestHandler:
except litellm.BudgetExceededError as e:
raise HTTPException(status_code=getattr(e, "status_code", 429), detail=str(e)) from None
except Exception as e: # noqa: BLE001 # untyped gate failure: retryable 503 for a DB outage, else fail closed 401
if PrismaDBExceptionHandler.is_database_service_unavailable_error(e):
raise HTTPException(
status_code=503,
detail="Service Unavailable, the authentication database is temporarily unreachable. Please retry shortly.",
) from None
MCPRequestHandler._raise_503_if_db_unavailable(e)
raise HTTPException(status_code=401, detail="Invalid or expired credential") from None
@staticmethod

View file

@ -5275,6 +5275,28 @@ class TestMCPDcrBridgeDelegateAdmission:
mapped = await self._enforce_with_gate_error(ConnectionError("could not reach database server"))
assert mapped.status_code == 503
async def test_db_outage_during_key_reload_surfaces_503_not_500(self):
"""A DB outage while reloading the admitted key surfaces a retryable 503, not the opaque 500 a
raw get_key_object transport error would otherwise propagate as, and not a 401 that masks the
outage as an auth failure. Regression for the reload-path exception gap."""
envelope = self._mint_bridge_envelope(key_hash=self._KEY_HASH)
scope = {
"type": "http",
"method": "POST",
"path": "/mcp/bridge_delegate_server",
"headers": [(b"authorization", f"Bearer {envelope}".encode("latin-1"))],
}
with (
patch("litellm.proxy._experimental.mcp_server.mcp_server_manager.global_mcp_server_manager") as mock_mgr,
patch("litellm.proxy.proxy_server.master_key", self._MASTER_KEY),
self._patch_key_reload(side_effect=ConnectionError("could not reach database server")),
):
mock_mgr.get_mcp_server_by_name.return_value = self._bridge_delegate_server()
with pytest.raises(HTTPException) as exc_info:
await MCPRequestHandler.process_mcp_request(scope)
assert exc_info.value.status_code == 503
async def test_blocked_state_bare_exception_stays_401(self):
"""A blocked team/project raises a bare Exception (no status) in common_checks, which the
standard pipeline renders as 401; the arm keeps failing those closed as 401, never a 500."""