fix(proxy): completion-marker ttl outliving hanging-request tracker entry

update_request_status wrote the "request completed" marker with
ttl = alerting_threshold + 100, while the hanging-request tracker entry
(HangingRequestCheck) lives for alerting_threshold * 1.5 +
HANGING_ALERT_BUFFER_TIME_SECONDS. Since the check loop only scans the
20 oldest tracked entries per pass, a sustained burst above ~8 req/min
can let a completed request's tracker entry sit in the backlog long
enough for its shorter-lived completion marker to expire first - the
check then finds no marker, assumes the request is still running, and
fires a false "hanging request" Slack alert for a request that already
finished in under a second.

Align the marker's ttl with the tracker entry's formula so a completed
request's marker always outlives the window in which it might still be
checked.

Fixes #43285

litellm/proxy/utils.py

@@ -1434,9 +1434,18 @@ class ProxyLogging:
         # current alerting threshold
         alerting_threshold: float = self.alerting_threshold
 
-        # add a 100 second buffer to the alerting threshold
-        # ensures we don't send errant hanging request slack alerts
-        alerting_threshold += 100
+        # This marker's ttl must outlive the hanging-request tracker entry
+        # (alerting_threshold * 1.5 + HANGING_ALERT_BUFFER_TIME_SECONDS, see
+        # HangingRequestCheck._get_metadata) - otherwise a completed request
+        # can have its marker expire before the tracker gets around to
+        # checking it (only the 20 oldest entries are scanned per pass),
+        # producing a false "hanging request" alert for a request that
+        # already finished successfully.
+        from litellm.types.integrations.slack_alerting import (
+            HANGING_ALERT_BUFFER_TIME_SECONDS,
+        )
+
+        alerting_threshold = alerting_threshold * 1.5 + HANGING_ALERT_BUFFER_TIME_SECONDS
 
         await self.internal_usage_cache.async_set_cache(
             key=f"request_status:{litellm_call_id}",

tests/test_litellm/proxy/test_proxy_utils.py

@@ -26,6 +26,35 @@ def test_get_custom_url(monkeypatch):
     assert custom_url == "http://0.0.0.0:4000/litellm/ui/"
 
 
+@pytest.mark.asyncio
+async def test_update_request_status_marker_ttl_outlives_tracker_entry():
+    """The completion marker written by update_request_status must have a
+    ttl >= the hanging-request tracker entry's ttl (alerting_threshold * 1.5
+    + HANGING_ALERT_BUFFER_TIME_SECONDS). Otherwise a completed request's
+    marker can expire before the hanging-request check gets around to
+    scanning it, producing a false "hanging request" alert. See #43285.
+    """
+    from litellm.types.integrations.slack_alerting import (
+        HANGING_ALERT_BUFFER_TIME_SECONDS,
+    )
+
+    proxy_logging_obj = ProxyLogging(user_api_key_cache=DualCache())
+    proxy_logging_obj.alerting = ["slack"]
+    proxy_logging_obj.alerting_threshold = 300
+
+    tracker_entry_ttl = 300 * 1.5 + HANGING_ALERT_BUFFER_TIME_SECONDS
+
+    with patch.object(
+        proxy_logging_obj.internal_usage_cache, "async_set_cache", new=AsyncMock()
+    ) as mock_set_cache:
+        await proxy_logging_obj.update_request_status(
+            litellm_call_id="test_call_id", status="success"
+        )
+
+    used_ttl = mock_set_cache.call_args.kwargs["ttl"]
+    assert used_ttl >= tracker_entry_ttl
+
+
 def test_proxy_only_error_true_for_llm_route():
     proxy_logging_obj = ProxyLogging(user_api_key_cache=DualCache())
     assert proxy_logging_obj._is_proxy_only_llm_api_error(