Merge dda896bbdc into 62f032cca5

litellm/proxy/_types.py

@@ -2526,6 +2526,10 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
         description="for `/models` endpoint, infers available model based on environment keys (e.g. OPENAI_API_KEY)",
     )
     background_health_checks: bool | None = Field(None, description="run health checks in background")
+    persist_background_health_check_results: bool = Field(
+        True,
+        description="persist background health check results to the database",
+    )
     health_check_interval: int = Field(300, description="background health check interval in seconds")
     health_check_concurrency: int | None = Field(
         None,

litellm/proxy/proxy_server.py

@@ -3580,9 +3580,10 @@ def _schedule_background_health_check_db_save(
     model_list: list,
     healthy_endpoints: list,
     unhealthy_endpoints: list,
+    persist_results: bool = True,
 ):
     """Fire-and-forget: persist health check results to DB if prisma is available."""
-    if prisma_client is None:
+    if prisma_client is None or not persist_results:
         return
     import time as time_module
 
@@ -3764,11 +3765,12 @@ async def _run_background_health_check():
         )
     background_health_check_loop_active = True
     verbose_proxy_logger.info(
-        "background_health_check_loop_started interval_seconds=%s max_concurrency=%s shared=%s details=%s thread_count=%d rss_mb=%s",
+        "background_health_check_loop_started interval_seconds=%s max_concurrency=%s shared=%s details=%s persist_results=%s thread_count=%d rss_mb=%s",
         health_check_interval,
         health_check_concurrency,
         use_shared_health_check,
         health_check_details,
+        general_settings.get("persist_background_health_check_results", True),
         threading.active_count(),
         _rss_mb_for_log(),
     )
@@ -3908,6 +3910,7 @@ async def _run_background_health_check():
             _llm_model_list,
             healthy_endpoints,
             unhealthy_endpoints,
+            persist_results=general_settings.get("persist_background_health_check_results", True),
         )
 
         # Write health state to router cache for health-check-driven routing

tests/test_litellm/proxy/proxy_server/test_background_health.py

@@ -19,14 +19,15 @@ from unittest.mock import AsyncMock, MagicMock
 
 import pytest
 
-import litellm.proxy.proxy_server as proxy_server
+from litellm.proxy import proxy_server
+from litellm.proxy._types import ConfigGeneralSettings
 from litellm.proxy.proxy_server import (
     _adaptive_router_flusher_loop,
     _get_endpoint_exception_status,
     _get_process_rss_mb,
+    _rss_mb_for_log,
     _run_background_health_check,
     _run_direct_health_check_with_instrumentation,
-    _rss_mb_for_log,
     _schedule_background_health_check_db_save,
     _write_health_state_to_router_cache,
 )
@@ -159,6 +160,18 @@ async def test_run_direct_health_check_raises_non_kwarg_typeerror(monkeypatch):
 # ---------------------------------------------------------------------------
 
 
+def test_background_health_check_db_persistence_config_defaults_enabled():
+    assert ConfigGeneralSettings().persist_background_health_check_results is True
+
+
+def test_background_health_check_db_persistence_config_can_be_disabled():
+    settings = ConfigGeneralSettings.model_validate(
+        {"persist_background_health_check_results": False}
+    )
+
+    assert settings.persist_background_health_check_results is False
+
+
 @pytest.mark.asyncio
 async def test_schedule_background_health_check_db_save_creates_task(monkeypatch):
     captured = {}
@@ -219,6 +232,24 @@ def test_schedule_background_health_check_db_save_noop_when_prisma_none():
     )
 
 
+def test_schedule_background_health_check_db_save_noop_when_persistence_disabled(
+    monkeypatch,
+):
+    create_task = MagicMock()
+    monkeypatch.setattr(asyncio, "create_task", create_task)
+
+    _schedule_background_health_check_db_save(
+        prisma_client=MagicMock(),
+        shared_health_manager=None,
+        model_list=[],
+        healthy_endpoints=[],
+        unhealthy_endpoints=[],
+        persist_results=False,
+    )
+
+    create_task.assert_not_called()
+
+
 @pytest.mark.asyncio
 async def test_schedule_background_health_check_db_save_invalid_no_event_loop_raises(
     monkeypatch,
@@ -527,6 +558,11 @@ async def test_run_background_health_check_runs_one_cycle_then_cancels(monkeypat
     monkeypatch.setattr(proxy_server, "use_shared_health_check", False)
     monkeypatch.setattr(proxy_server, "redis_usage_cache", None)
     monkeypatch.setattr(proxy_server, "prisma_client", None)
+    monkeypatch.setattr(
+        proxy_server,
+        "general_settings",
+        {"persist_background_health_check_results": False},
+    )
     monkeypatch.setattr(proxy_server, "background_health_check_loop_active", False)
     monkeypatch.setattr(
         proxy_server,
@@ -547,8 +583,9 @@ async def test_run_background_health_check_runs_one_cycle_then_cancels(monkeypat
         "_run_direct_health_check_with_instrumentation",
         _fake_direct,
     )
+    scheduled = MagicMock()
     monkeypatch.setattr(
-        proxy_server, "_schedule_background_health_check_db_save", lambda *a, **kw: None
+        proxy_server, "_schedule_background_health_check_db_save", scheduled
     )
     monkeypatch.setattr(
         proxy_server, "_write_health_state_to_router_cache", lambda *a, **kw: None
@@ -574,11 +611,13 @@ async def test_run_background_health_check_runs_one_cycle_then_cancels(monkeypat
         {
             "healthy_count": proxy_server.health_check_results["healthy_count"],
             "unhealthy_count": proxy_server.health_check_results["unhealthy_count"],
+            "persist_results": scheduled.call_args.kwargs["persist_results"],
             "sleep_invoked": sleep_calls["n"] >= 1,
         }
     ) == {
         "healthy_count": 1,
         "unhealthy_count": 1,
+        "persist_results": False,
         "sleep_invoked": True,
     }
 

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

@@ -25663,6 +25663,12 @@ export interface components {
              * @description Default upstream request timeout in seconds for native and custom pass-through endpoints that use pass_through_request. Defaults to 600 when unset.
              */
             pass_through_request_timeout?: number | null;
+            /**
+             * Persist Background Health Check Results
+             * @description persist background health check results to the database
+             * @default true
+             */
+            persist_background_health_check_results: boolean;
             /**
              * Plugins
              * @description external services registered as embeddable UI plugins