fix(alerts): repair detector tests and tracing

litellm/integrations/otel/model/spans.py

@@ -261,6 +261,7 @@ _PRISMA_VIEWS: Final[frozenset[str]] = frozenset(
 _PRISMA_MODELS: Final[frozenset[str]] = frozenset(
     (
         "LiteLLM_Alert",
+        "LiteLLM_PerformanceAlertState",
         "LiteLLM_BudgetTable",
         "LiteLLM_CredentialsTable",
         "LiteLLM_ProxyModelTable",

litellm/proxy/proxy_server.py

@@ -11877,7 +11877,25 @@ class ProxyStartupEvent:
             read_general_settings=_current_general_settings,
         )
 
-        cls._initialize_performance_alerts_job(scheduler=scheduler, prisma_client=prisma_client)
+        from litellm.proxy.spend_tracking.performance_alerts import RUN_INTERVAL, run_performance_alerts
+        from litellm.repositories.performance_alert_repository import PerformanceAlertRepository
+
+        performance_alert_repository: Final = PerformanceAlertRepository(prisma_client)
+
+        async def run_performance_alerts_job() -> None:
+            if _current_general_settings().get("performance_alerts_enabled") is True:
+                await run_performance_alerts(performance_alert_repository)
+
+        scheduler.add_job(  # pyright: ignore[reportUnknownMemberType]  # APScheduler does not provide typed job methods
+            run_performance_alerts_job,
+            "interval",
+            seconds=RUN_INTERVAL.total_seconds(),
+            id="performance_alerts",
+            replace_existing=True,
+            max_instances=1,
+            coalesce=True,
+            misfire_grace_time=APSCHEDULER_MISFIRE_GRACE_TIME,
+        )
 
         ### PTU DAILY ROLLUP ###
         from litellm.proxy.spend_tracking.ptu_feature_flag import (
@@ -12255,28 +12273,6 @@ class ProxyStartupEvent:
                 "LITELLM_EXPIRED_UI_SESSION_KEY_CLEANUP_ENABLED=true to enable)"
             )
 
-    @classmethod
-    def _initialize_performance_alerts_job(cls, scheduler: AsyncIOScheduler, prisma_client: PrismaClient) -> None:
-        from litellm.proxy.spend_tracking.performance_alerts import RUN_INTERVAL, run_performance_alerts
-        from litellm.repositories.performance_alert_repository import PerformanceAlertRepository
-
-        repository: Final = PerformanceAlertRepository(prisma_client)
-
-        async def run() -> None:
-            if _current_general_settings().get("performance_alerts_enabled") is True:
-                await run_performance_alerts(repository)
-
-        scheduler.add_job(  # pyright: ignore[reportUnknownMemberType]  # APScheduler does not provide typed job methods
-            run,
-            "interval",
-            seconds=RUN_INTERVAL.total_seconds(),
-            id="performance_alerts",
-            replace_existing=True,
-            max_instances=1,
-            coalesce=True,
-            misfire_grace_time=APSCHEDULER_MISFIRE_GRACE_TIME,
-        )
-
     @classmethod
     def _initialize_spend_capture_rate_check_job(
         cls,

tests/unit/proxy/management_endpoints/test_model_insights_endpoints.py

@@ -253,6 +253,7 @@ async def test_model_insights_reads_back_what_the_rollup_wrote() -> None:
         "request_tags": '["task:debugging"]',
         "custom_llm_provider": "openai",
         "status": "success",
+        "request_duration_ms": None,
     }
 
     transactions = (

tests/unit/proxy/test_proxy_server_endpoints_and_startup.py

@@ -1656,6 +1656,7 @@ class MockPrisma:
         self.database_url = database_url
         self.proxy_logging_obj = proxy_logging_obj
         self.http_client = http_client
+        self.writer_db = MagicMock()
 
     async def connect(self):
         pass

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -31397,6 +31397,12 @@ export interface components {
              * @description If True (default), a locally-managed user's password must contain an uppercase letter.
              */
             password_policy_require_uppercase?: boolean | null;
+            /**
+             * Performance Alerts Enabled
+             * @description Check yesterday's usage for the gateway, models, teams, users and keys against the week before, and save unusual spend, traffic, response time and failure rate changes as alerts. Off by default.
+             * @default false
+             */
+            performance_alerts_enabled: boolean;
             /**
              * Plugins
              * @description external services registered as embeddable UI plugins