Merge pull request #3599 from taralika/patch-1

Ignore 0 failures and 0s latency in daily slack reports

litellm/integrations/slack_alerting.py

@@ -347,8 +347,9 @@ class SlackAlerting(CustomLogger):
 
         all_none = True
         for val in combined_metrics_values:
-            if val is not None:
+            if val is not None and val > 0:
                 all_none = False
+                break
 
         if all_none:
             return False
@@ -366,12 +367,13 @@ class SlackAlerting(CustomLogger):
             for value in failed_request_values
         ]
 
-        ## Get the indices of top 5 keys with the highest numerical values (ignoring None values)
+        ## Get the indices of top 5 keys with the highest numerical values (ignoring None and 0 values)
         top_5_failed = sorted(
             range(len(replaced_failed_values)),
             key=lambda i: replaced_failed_values[i],
             reverse=True,
         )[:5]
+        top_5_failed = [index for index in top_5_failed if replaced_failed_values[index] > 0]
         
         # find top 5 slowest
         # Replace None values with a placeholder value (-1 in this case)
@@ -381,17 +383,20 @@ class SlackAlerting(CustomLogger):
             for value in latency_values
         ]
 
-        # Get the indices of top 5 values with the highest numerical values (ignoring None values)
+        # Get the indices of top 5 values with the highest numerical values (ignoring None and 0 values)
         top_5_slowest = sorted(
             range(len(replaced_slowest_values)),
             key=lambda i: replaced_slowest_values[i],
             reverse=True,
         )[:5]
+        top_5_slowest = [index for index in top_5_slowest if replaced_slowest_values[index] > 0]
 
         # format alert -> return the litellm model name + api base
         message = f"\n\nHere are today's key metrics 📈: \n\n"
 
-        message += "\n\n*❗️ Top 5 Deployments with Most Failed Requests:*\n\n"
+        message += "\n\n*❗️ Top Deployments with Most Failed Requests:*\n\n"
+        if not top_5_failed:
+            message += "\tNone\n"
         for i in range(len(top_5_failed)):
             key = failed_request_keys[top_5_failed[i]].split(":")[0]
             _deployment = router.get_model_info(key)
@@ -411,7 +416,9 @@ class SlackAlerting(CustomLogger):
             value = replaced_failed_values[top_5_failed[i]]
             message += f"\t{i+1}. Deployment: `{deployment_name}`, Failed Requests: `{value}`,  API Base: `{api_base}`\n"
 
-        message += "\n\n*😅 Top 5 Slowest Deployments:*\n\n"
+        message += "\n\n*😅 Top Slowest Deployments:*\n\n"
+        if not top_5_slowest:
+            message += "\tNone\n"
         for i in range(len(top_5_slowest)):
             key = latency_keys[top_5_slowest[i]].split(":")[0]
             _deployment = router.get_model_info(key)

litellm/tests/test_alerting.py

@@ -359,3 +359,49 @@ async def test_send_llm_exception_to_slack():
     )
 
     await asyncio.sleep(3)
+
+
+# test models with 0 metrics are ignored
+@pytest.mark.asyncio
+async def test_send_daily_reports_ignores_zero_values():
+    router = MagicMock()
+    router.get_model_ids.return_value = ['model1', 'model2', 'model3']
+    
+    slack_alerting = SlackAlerting(internal_usage_cache=MagicMock())
+    # model1:failed=None, model2:failed=0, model3:failed=10, model1:latency=0; model2:latency=0; model3:latency=None
+    slack_alerting.internal_usage_cache.async_batch_get_cache = AsyncMock(return_value=[None, 0, 10, 0, 0, None])
+    slack_alerting.internal_usage_cache.async_batch_set_cache = AsyncMock()
+
+    router.get_model_info.side_effect = lambda x: {"litellm_params": {"model": x}}
+    
+    with patch.object(slack_alerting, 'send_alert', new=AsyncMock()) as mock_send_alert:
+        result = await slack_alerting.send_daily_reports(router)
+        
+        # Check that the send_alert method was called
+        mock_send_alert.assert_called_once()
+        message = mock_send_alert.call_args[1]['message']
+        
+        # Ensure the message includes only the non-zero, non-None metrics
+        assert "model3" in message
+        assert "model2" not in message
+        assert "model1" not in message
+    
+    assert result == True
+
+
+# test no alert is sent if all None or 0 metrics
+@pytest.mark.asyncio
+async def test_send_daily_reports_all_zero_or_none():
+    router = MagicMock()
+    router.get_model_ids.return_value = ['model1', 'model2', 'model3']
+    
+    slack_alerting = SlackAlerting(internal_usage_cache=MagicMock())
+    slack_alerting.internal_usage_cache.async_batch_get_cache = AsyncMock(return_value=[None, 0, None, 0, None, 0])
+    
+    with patch.object(slack_alerting, 'send_alert', new=AsyncMock()) as mock_send_alert:
+        result = await slack_alerting.send_daily_reports(router)
+        
+        # Check that the send_alert method was not called
+        mock_send_alert.assert_not_called()
+    
+    assert result == False