fix(proxy): share single _ProxyDBLogger instance across callback lists

cost_tracking() was creating two separate _ProxyDBLogger instances — one
for sync callbacks and one for async. This doubles DB connection pressure
and causes state divergence between the two loggers. Reuse a single
instance for both callback registrations.

litellm/proxy/proxy_server.py

@@ -1892,10 +1892,9 @@ def load_from_azure_key_vault(use_azure_key_vault: bool = False):
 def cost_tracking():
     global prisma_client
     if prisma_client is not None:
-        litellm.logging_callback_manager.add_litellm_callback(_ProxyDBLogger())
-        litellm.logging_callback_manager.add_litellm_async_success_callback(
-            _ProxyDBLogger()
-        )
+        db_logger = _ProxyDBLogger()
+        litellm.logging_callback_manager.add_litellm_callback(db_logger)
+        litellm.logging_callback_manager.add_litellm_async_success_callback(db_logger)
 
 
 async def get_current_spend(counter_key: str, fallback_spend: float) -> float:
@@ -2710,11 +2709,9 @@ def run_ollama_serve():
         with open(os.devnull, "w") as devnull:
             subprocess.Popen(command, stdout=devnull, stderr=devnull)
     except Exception as e:
-        verbose_proxy_logger.debug(
-            f"""
+        verbose_proxy_logger.debug(f"""
             LiteLLM Warning: proxy started with `ollama` model\n`ollama serve` failed with Exception{e}. \nEnsure you run `ollama serve`
-        """
-        )
+        """)
 
 
 def _get_process_rss_mb() -> Optional[float]:

tests/test_litellm/proxy/test_proxy_server.py

@@ -7487,3 +7487,36 @@ class TestSortModelsByDisplayName:
             all_models=models, sort_by="model_name", sort_order="asc"
         )
         assert [m["model_name"] for m in sorted_models] == ["alpha", "beta"]
+
+
+class TestCostTrackingSingleDBLogger:
+    def test_cost_tracking_creates_single_db_logger_instance(self):
+        from unittest.mock import MagicMock, patch
+
+        from litellm.proxy.proxy_server import cost_tracking
+
+        mock_callback_manager = MagicMock()
+
+        with patch("litellm.proxy.proxy_server.prisma_client", new="fake-client"):
+            with patch("litellm.logging_callback_manager", mock_callback_manager):
+                cost_tracking()
+
+        add_callback_call = mock_callback_manager.add_litellm_callback.call_args[0][0]
+        add_async_call = (
+            mock_callback_manager.add_litellm_async_success_callback.call_args[0][0]
+        )
+        assert add_callback_call is add_async_call
+
+    def test_cost_tracking_noop_when_prisma_client_is_none(self):
+        from unittest.mock import MagicMock, patch
+
+        from litellm.proxy.proxy_server import cost_tracking
+
+        mock_callback_manager = MagicMock()
+
+        with patch("litellm.proxy.proxy_server.prisma_client", new=None):
+            with patch("litellm.logging_callback_manager", mock_callback_manager):
+                cost_tracking()
+
+        mock_callback_manager.add_litellm_callback.assert_not_called()
+        mock_callback_manager.add_litellm_async_success_callback.assert_not_called()