Fix/background health check (#10887)

* fix: improve health check logic by deep copying model list on each iteration

* test: add async test for background health check reflecting model list changes

* fix: validate health check interval before executing background health check

* fix: specify type for health check results dictionary

litellm/proxy/proxy_server.py

@@ -835,7 +835,7 @@ use_background_health_checks = None
 use_queue = False
 health_check_interval = None
 health_check_details = None
-health_check_results = {}
+health_check_results: dict[str, Union[int, List[Dict[str, Any]]]] = {}
 queue: List = []
 litellm_proxy_budget_name = "litellm-proxy-budget"
 litellm_proxy_admin_name = LITELLM_PROXY_ADMIN_NAME
@@ -1231,13 +1231,15 @@ async def _run_background_health_check():
     """
     global health_check_results, llm_model_list, health_check_interval, health_check_details
 
-    # make 1 deep copy of llm_model_list -> use this for all background health checks
-    _llm_model_list = copy.deepcopy(llm_model_list)
-
-    if _llm_model_list is None:
+    if health_check_interval is None or not isinstance(
+            health_check_interval, int
+        ) or health_check_interval <= 0:
         return
 
     while True:
+        # make 1 deep copy of llm_model_list on every health check iteration
+        _llm_model_list = copy.deepcopy(llm_model_list) or []
+
         healthy_endpoints, unhealthy_endpoints = await perform_health_check(
             model_list=_llm_model_list, details=health_check_details
         )
@@ -1248,10 +1250,7 @@ async def _run_background_health_check():
         health_check_results["healthy_count"] = len(healthy_endpoints)
         health_check_results["unhealthy_count"] = len(unhealthy_endpoints)
 
-        if health_check_interval is not None and isinstance(
-            health_check_interval, float
-        ):
-            await asyncio.sleep(health_check_interval)
+        await asyncio.sleep(health_check_interval)
 
 
 class StreamingCallbackError(Exception):

tests/proxy_unit_tests/test_proxy_server.py

@@ -2194,6 +2194,50 @@ async def test_get_ui_settings_spend_logs_threshold():
     proxy_state.set_proxy_state_variable("spend_logs_row_count", 0)
 
 
+@pytest.mark.asyncio
+async def test_run_background_health_check_reflects_llm_model_list(monkeypatch):
+    """
+    Test that _run_background_health_check reflects changes to llm_model_list in each health check iteration.
+    """
+    import litellm.proxy.proxy_server as proxy_server
+    import copy
+
+    test_model_list_1 = [{"model_name": "model-a"}]
+    test_model_list_2 = [{"model_name": "model-b"}]
+    called_model_lists = []
+
+    async def fake_perform_health_check(model_list, details):
+        called_model_lists.append(copy.deepcopy(model_list))
+        return (["healthy"], ["unhealthy"])
+
+    monkeypatch.setattr(proxy_server, "health_check_interval", 1)
+    monkeypatch.setattr(proxy_server, "health_check_details", None)
+    monkeypatch.setattr(proxy_server, "llm_model_list", copy.deepcopy(test_model_list_1))
+    monkeypatch.setattr(proxy_server, "perform_health_check", fake_perform_health_check)
+    monkeypatch.setattr(proxy_server, "health_check_results", {})
+
+    async def fake_sleep(interval):
+        raise asyncio.CancelledError()
+
+    monkeypatch.setattr(asyncio, "sleep", fake_sleep)
+
+    try:
+        await proxy_server._run_background_health_check()
+    except asyncio.CancelledError:
+        pass
+
+    monkeypatch.setattr(proxy_server, "llm_model_list", copy.deepcopy(test_model_list_2))
+
+    try:
+        await proxy_server._run_background_health_check()
+    except asyncio.CancelledError:
+        pass
+
+    assert len(called_model_lists) >= 2
+    assert called_model_lists[0] == test_model_list_1
+    assert called_model_lists[1] == test_model_list_2
+
+
 def test_get_timeout_from_request():
     from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup