fix(otel): use attempt start for deferred retry failures

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
shivam 2026-11-14 02:31:33 +00:00
parent b4161834a6
commit 11bdd8d4a3
4 changed files with 41 additions and 7 deletions

View file

@ -2983,8 +2983,11 @@ class Logging(LiteLLMLoggingBaseClass):
},
}
attempt.litellm_params = attempt.model_call_details["litellm_params"]
attempt_start_time: Final = (
marker if start_time is None and isinstance(marker, datetime.datetime) else start_time
)
attempt_start, attempt_end = attempt._failure_handler_helper_fn(
exception, traceback_exception, start_time, end_time, estimate_prompt_tokens=False
exception, traceback_exception, attempt_start_time, end_time, estimate_prompt_tokens=False
)
if attempt._should_estimate_failed_prompt():
attempt.model_call_details[

View file

@ -47,6 +47,7 @@ class Span(TypedDict):
attributes: ReadOnly[Mapping[str, JsonValue]]
resource: ReadOnly[Mapping[str, JsonValue]]
status_code: ReadOnly[NotRequired[int]]
start_time_unix_nano: ReadOnly[NotRequired[int]]
class _SpanListing(TypedDict):
@ -89,6 +90,7 @@ def _proto_spans(body: bytes) -> list[Span]:
parent_span_id=span.parent_span_id.hex(),
kind=span.kind,
status_code=span.status.code,
start_time_unix_nano=span.start_time_unix_nano,
name=span.name,
attributes={attribute.key: scalar(attribute.value) for attribute in span.attributes},
resource={attribute.key: scalar(attribute.value) for attribute in resource.resource.attributes},
@ -121,6 +123,7 @@ def _json_spans(body: bytes) -> list[Span]:
parent_span_id=str(span.get("parentSpanId", "")),
kind=int(span.get("kind", 0)),
status_code=int(span.get("status", {}).get("code", 0)),
start_time_unix_nano=int(span.get("startTimeUnixNano", 0)),
name=str(span.get("name", "")),
attributes={attribute["key"]: scalar(attribute.get("value")) for attribute in span.get("attributes", [])},
resource={

View file

@ -224,6 +224,10 @@ def test_failed_retry_stream_has_one_span_per_upstream_attempt(retry_rig: RetryR
)
assert len(spans) == 5, spans
assert all(span.get("status_code") == 2 for span in spans), spans
assert all(span.get("start_time_unix_nano", 0) > 0 for span in spans), spans
assert len({span.get("start_time_unix_nano") for span in spans}) == 5, spans
last_attempt: Final = max(spans, key=lambda span: span.get("start_time_unix_nano", 0))
assert int(str(last_attempt["attributes"].get("server.port"))) == urlsplit(retry_rig.broken_stream.url).port, spans
ports: Final = sorted(int(str(span["attributes"].get("server.port"))) for span in spans)
assert ports == sorted([urlsplit(retry_rig.primary.url).port] * 4 + [urlsplit(retry_rig.broken_stream.url).port]), (
spans

View file

@ -152,6 +152,7 @@ class _FailureAttemptRecorder(CustomLogger):
def __init__(self) -> None:
super().__init__()
self.payloads: list[StandardLoggingPayload] = []
self.failure_starts: list[datetime.datetime] = []
self.successes = 0
async def async_log_success_event(
@ -171,6 +172,7 @@ class _FailureAttemptRecorder(CustomLogger):
end_time: datetime.datetime,
) -> None:
self.payloads.append(cast(StandardLoggingPayload, kwargs["standard_logging_object"]))
self.failure_starts.append(start_time)
class _EveryFailureAttemptRecorder(_FailureAttemptRecorder):
@ -179,9 +181,11 @@ class _EveryFailureAttemptRecorder(_FailureAttemptRecorder):
@pytest.mark.asyncio
@pytest.mark.parametrize("dynamic", [True, False])
@pytest.mark.parametrize("omit_start_time", [False, True])
async def test_reused_logging_dispatches_only_opted_in_loggers_for_each_failed_attempt(
monkeypatch: pytest.MonkeyPatch,
dynamic: bool,
omit_start_time: bool,
) -> None:
normal: Final = _FailureAttemptRecorder()
telemetry: Final = _EveryFailureAttemptRecorder()
@ -209,14 +213,18 @@ async def test_reused_logging_dispatches_only_opted_in_loggers_for_each_failed_a
logging_obj.model_call_details["api_call_start_time"] = marker
logging_obj.model_call_details["current_attempt_api_call_start_time"] = marker
prior_payload: Final = logging_obj.model_call_details.get("standard_logging_object")
await logging_obj.async_failure_handler(error, "trace", marker, marker + datetime.timedelta(seconds=1))
await logging_obj.async_failure_handler(
error, "trace", None if omit_start_time else marker, marker + datetime.timedelta(seconds=1)
)
if index > 0:
assert logging_obj.model_call_details["standard_logging_object"] is prior_payload
assert logging_obj.model_call_details["async_failure_logged_attempt"] == start + datetime.timedelta(
seconds=1
)
saved_payload: Final = logging_obj.model_call_details["standard_logging_object"]
await logging_obj.async_failure_handler(error, "trace", marker, marker + datetime.timedelta(seconds=1))
await logging_obj.async_failure_handler(
error, "trace", None if omit_start_time else marker, marker + datetime.timedelta(seconds=1)
)
assert logging_obj.model_call_details["standard_logging_object"] is saved_payload
await logging_obj.async_success_handler(
litellm.ModelResponse(model="gpt-4o"), start, start + datetime.timedelta(seconds=5)
@ -224,6 +232,13 @@ async def test_reused_logging_dispatches_only_opted_in_loggers_for_each_failed_a
assert [p["api_base"] for p in telemetry.payloads] == [f"https://region-{i}.example/v1" for i in range(3)]
assert [p["api_base"] for p in normal.payloads] == ["https://region-0.example/v1"]
assert all(p["status"] == "failure" for p in telemetry.payloads)
first_start: Final = start if omit_start_time else start + datetime.timedelta(seconds=1)
assert normal.failure_starts == [first_start]
assert telemetry.failure_starts == [
first_start,
start + datetime.timedelta(seconds=2),
start + datetime.timedelta(seconds=3),
]
assert normal.successes == telemetry.successes == 1
@ -251,14 +266,21 @@ async def test_retry_spans_preserve_request_failure_metric_count(monkeypatch: py
monkeypatch.setattr(litellm, "_async_failure_callback", [telemetry])
start: Final = datetime.datetime(2026, 1, 1)
logging_obj: Final = Logging(
model="gpt-4o", messages=[{"role": "user", "content": "hello"}], stream=False,
call_type="acompletion", start_time=start, litellm_call_id="metric-attempts", function_id="fn",
model="gpt-4o",
messages=[{"role": "user", "content": "hello"}],
stream=False,
call_type="acompletion",
start_time=start,
litellm_call_id="metric-attempts",
function_id="fn",
)
error: Final = litellm.ServiceUnavailableError("outage", model="gpt-4o", llm_provider="openai")
for index in range(3):
marker: Final = start + datetime.timedelta(seconds=index + 1)
logging_obj.model_call_details["litellm_params"] = {
"api_base": f"https://region-{index}.example/v1", "custom_llm_provider": "openai", "metadata": {},
"api_base": f"https://region-{index}.example/v1",
"custom_llm_provider": "openai",
"metadata": {},
}
logging_obj.model_call_details["api_call_start_time"] = marker
logging_obj.model_call_details["current_attempt_api_call_start_time"] = marker
@ -269,7 +291,9 @@ async def test_retry_spans_preserve_request_failure_metric_count(monkeypatch: py
assert metrics is not None
for resource in metrics.resource_metrics:
for scope in resource.scope_metrics:
duration: Final = next(metric for metric in scope.metrics if metric.name == "gen_ai.client.operation.duration")
duration: Final = next(
metric for metric in scope.metrics if metric.name == "gen_ai.client.operation.duration"
)
assert sum(point.count for point in duration.data.data_points) == 1