fix(langfuse): avoid invalid v4 observation timing

litellm/integrations/langfuse/langfuse.py

@@ -695,9 +695,6 @@ class LangFuseLogger:
             if output is not None and isinstance(output, str) and level == "ERROR":
                 generation_params["status_message"] = output
 
-            if self._supports_completion_start_time():
-                generation_params["completion_start_time"] = kwargs.get("completion_start_time", None)
-
             from langfuse import propagate_attributes
 
             trace_id_value = trace_params.get("id")
@@ -723,8 +720,6 @@ class LangFuseLogger:
             propagated_trace_name = (
                 trace_name_value if isinstance(trace_name_value, str) and trace_name_value else trace_name
             )
-            end_time_ns = int(end_time.timestamp() * 1_000_000_000) if end_time is not None else None
-
             with propagate_attributes(
                 user_id=propagated_user_id,
                 session_id=propagated_session_id,
@@ -749,8 +744,8 @@ class LangFuseLogger:
                     standard_logging_object=standard_logging_object,
                 )
                 generation_client = trace.start_observation(**generation_params)
-                generation_client.end(end_time=end_time_ns)
-                trace.end(end_time=end_time_ns)
+                generation_client.end()
+                trace.end()
 
             return trace_context["trace_id"], generation_client.id
         except Exception:

tests/logging_callback_tests/test_langfuse_unit_tests.py

@@ -1,5 +1,6 @@
 import os
 import sys
+from datetime import datetime, timedelta
 
 sys.path.insert(0, os.path.abspath("../.."))  # Adds the parent directory to the system-path
 
@@ -339,6 +340,45 @@ def test_langfuse_v4_observations_propagate_trace_attributes():
         assert span.attributes["langfuse.trace.metadata.request_type"] == "completion"
 
 
+def test_langfuse_v4_observations_do_not_use_historical_end_times():
+    span_exporter = InMemorySpanExporter()
+    langfuse_client = Langfuse(
+        public_key="test-timing-public-key",
+        secret_key="test-secret-key",
+        host="https://cloud.langfuse.com",
+        span_exporter=span_exporter,
+    )
+    logger = object.__new__(LangFuseLogger)
+    logger.Langfuse = langfuse_client
+    logger.langfuse_sdk_version = "4.7.0"
+    end_time = datetime.now() - timedelta(seconds=10)
+
+    logger._log_langfuse_v2(
+        user_id=None,
+        metadata={"trace_id": "a" * 32},
+        litellm_params={"metadata": {}},
+        output={"role": "assistant", "content": "Hello"},
+        start_time=end_time - timedelta(seconds=1),
+        end_time=end_time,
+        kwargs={
+            "model": "openai/test-model",
+            "call_type": "completion",
+            "completion_start_time": end_time - timedelta(milliseconds=500),
+            "standard_logging_object": None,
+        },
+        optional_params={},
+        input={"messages": [{"role": "user", "content": "Hello"}]},
+        response_obj=None,
+        level="DEFAULT",
+        litellm_call_id="call-123",
+    )
+
+    langfuse_client.flush()
+    spans = span_exporter.get_finished_spans()
+    assert len(spans) == 2
+    assert all(span.end_time >= span.start_time for span in spans)
+
+
 def test_get_chat_content_for_langfuse():
     """
     Test that _get_chat_content_for_langfuse correctly extracts content from chat completion responses