fix(logging): record the exception's own frames on OTel and MLflow when a redacted failure logged no traceback

litellm/integrations/mlflow.py

@@ -5,7 +5,7 @@ from typing import TYPE_CHECKING, Any, Final
 
 from litellm._logging import verbose_logger
 from litellm.integrations.custom_logger import CustomLogger
-from litellm.litellm_core_utils.redact_messages import scrubbed_exception_text
+from litellm.litellm_core_utils.redact_messages import scrubbed_exception_record
 
 if TYPE_CHECKING:
     from litellm.types.utils import StandardLoggingPayload
@@ -90,7 +90,7 @@ class MlflowLogger(CustomLogger):
 
             # Record exception info as event
             if exception := kwargs.get("exception"):
-                scrubbed: Final = scrubbed_exception_text(kwargs)
+                scrubbed: Final = scrubbed_exception_record(kwargs)
                 span.add_event(
                     SpanEvent.from_exception(exception)
                     if scrubbed is None

litellm/integrations/opentelemetry.py

@@ -36,7 +36,7 @@ from litellm.integrations.otel.model.semconv import LiteLLM, Metric
 from litellm.integrations.otel.plumbing.otlp_tls import resolve_otlp_http_tls
 from litellm.integrations.otel.routing import routing_decision_attributes
 from litellm.litellm_core_utils.internal_call_metadata import is_unbilled_non_inference_call_from_params
-from litellm.litellm_core_utils.redact_messages import scrubbed_exception_text
+from litellm.litellm_core_utils.redact_messages import scrubbed_exception_record
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.litellm_core_utils.secret_redaction import redact_string
 from litellm.litellm_core_utils.service_tier_utils import (
@@ -2272,7 +2272,7 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
 
             # Get the exception object if available
             exception: Final = kwargs.get("exception")
-            scrubbed: Final = scrubbed_exception_text(kwargs)
+            scrubbed: Final = scrubbed_exception_record(kwargs)
 
             if exception is not None and scrubbed is not None:
                 span.record_exception(

tests/unit/integrations/test_mlflow.py

@@ -2,6 +2,7 @@ import asyncio
 import json
 import threading
 from datetime import datetime
+from typing import Final
 from unittest.mock import MagicMock, patch
 
 # Adds the grandparent directory to sys.path to allow importing project modules
@@ -321,3 +322,42 @@ def test_mlflow_failure_event_records_the_error_text_the_payload_logged(scrubbed
         assert event.attributes["exception.type"] == "ValueError"
         assert event.attributes["exception.message"] == error_message
         assert event.attributes["exception.stacktrace"] == ("frames" if scrubbed else "raw")
+
+
+def _raise_provider_error(message: str) -> None:
+    raise ValueError(message)
+
+
+def test_mlflow_failure_event_records_the_exceptions_own_frames_when_no_traceback_was_logged() -> None:
+    secret: Final = "my ssn is 123-45-6789"
+    modules: Final = _mock_mlflow_modules()
+    modules["mlflow"].get_current_active_span.return_value = None
+    modules["mlflow.entities"].SpanEvent = _ExceptionSpanEvent
+    with pytest.raises(ValueError, match="bad input") as raised:
+        _raise_provider_error(f"bad input: {secret}")
+    with patch.dict("sys.modules", modules):
+        from litellm.integrations.mlflow import MlflowLogger
+
+        mlflow_logger: Final = MlflowLogger()
+        span: Final = MagicMock(parent_id=None, request_id="req-1")
+        mlflow_logger._client = MagicMock(**{"start_trace.return_value": span})
+
+        mlflow_logger._handle_failure(
+            kwargs={
+                "litellm_call_id": "x",
+                "messages": [{"role": "user", "content": "hi"}],
+                "exception": raised.value,
+                "standard_logging_object": {
+                    "request_tags": [],
+                    "error_information": {"error_message": "bad input: redacted-by-litellm", "traceback": ""},
+                },
+            },
+            response_obj=None,
+            start_time=datetime(2025, 1, 1),
+            end_time=datetime(2025, 1, 1),
+        )
+
+    stacktrace: Final = span.add_event.call_args.args[0].attributes["exception.stacktrace"]
+    assert secret not in stacktrace
+    assert "in _raise_provider_error" in stacktrace
+    assert stacktrace.endswith("ValueError: bad input: redacted-by-litellm\n")

tests/unit/integrations/test_opentelemetry.py

@@ -7672,3 +7672,40 @@ def test_record_exception_on_span_records_the_scrubbed_error_text_with_the_same_
     assert event.attributes["exception.stacktrace"] == "frames"
     assert event.attributes["exception.type"] == "ValueError"
     assert event.attributes["exception.escaped"] == "False"
+
+
+def _raise_provider_error(message: str) -> None:
+    raise ValueError(message)
+
+
+def test_record_exception_on_span_records_the_exceptions_own_frames_when_no_traceback_was_logged() -> None:
+    from opentelemetry.sdk.trace import TracerProvider
+    from opentelemetry.sdk.trace.export import SimpleSpanProcessor
+    from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
+
+    from litellm.integrations.opentelemetry import OpenTelemetry
+
+    exporter: Final = InMemorySpanExporter()
+    provider: Final = TracerProvider()
+    provider.add_span_processor(SimpleSpanProcessor(exporter))
+    secret: Final = "my ssn is 123-45-6789"
+    with pytest.raises(ValueError, match="bad input") as raised:
+        _raise_provider_error(f"bad input: {secret}")
+
+    with provider.get_tracer(__name__).start_as_current_span("failure") as span:
+        OpenTelemetry()._record_exception_on_span(
+            span=span,
+            kwargs={
+                "exception": raised.value,
+                "standard_logging_object": {
+                    "error_information": {"error_message": "bad input: redacted-by-litellm", "traceback": ""}
+                },
+            },
+        )
+
+    (event,) = exporter.get_finished_spans()[0].events
+    stacktrace: Final = str(event.attributes["exception.stacktrace"])
+    assert secret not in str(dict(event.attributes))
+    assert stacktrace.startswith("Traceback (most recent call last):\n")
+    assert "in _raise_provider_error" in stacktrace
+    assert stacktrace.endswith("ValueError: bad input: redacted-by-litellm\n")