Same agent Smaller bill. The tools and infrastructure to get more work from your AI. Fewer tokens. Less compute. Lower cost.
TypeScript agent-harnessagent-harness-engineeringai-engineeringai-observabilityamd-gpuclaude-codeclickhousecodexcoding-agentcoding-agentscursorgenaigpu-monitoringllm-observabilityllmopsnvidia-smiobservabilityopen-sourceopenaiopentelemetry
# pylint: disable=protected-access, missing-function-docstring, duplicate-code, too-few-public-methods
"""Async runtime coverage for the `gen_ai.client.operation.duration` unit.
`test_operation_duration_unit.py` asserts the unit on the synchronous wrappers and
adds a static guard over the sources. Five of the nine original sites are in the
async modules, whose runtime behaviour nothing exercises: the async paths record
the attribute from three different places, and a static guard cannot tell which
`* 1000` is the semconv attribute and which is openlit's internal millisecond
bookkeeping.
These tests await the real async wrapper with a frozen clock and assert the
recorded attribute is elapsed **seconds** on both shapes:
- a normal awaited call,
- a call whose wrapped function raises.
"""
import asyncio
from unittest.mock import patch
import pytest
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import (
InMemorySpanExporter,
)
import openlit
from openlit._config import OpenlitConfig
from openlit.instrumentation.crawl4ai import async_crawl4ai as async_mod
from openlit.semcov import SemanticConvention
# Initialise OpenLIT the way the rest of this suite does, so class-level
# configuration such as custom_span_attributes is populated.
openlit.init(
environment="openlit-python-testing",
application_name="openlit-python-async-duration-test",
)
DURATION = SemanticConvention.GEN_AI_CLIENT_OPERATION_DURATION
ELAPSED_SECONDS = 2.5
@pytest.fixture(autouse=True)
def _reset_config():
OpenlitConfig.reset_to_defaults()
yield
OpenlitConfig.reset_to_defaults()
def _tracer_with_exporter():
exporter = InMemorySpanExporter()
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
return provider.get_tracer(__name__), exporter
def _recorded_durations(exporter):
"""Every value the finished spans carry for the duration attribute."""
return [
span.attributes[DURATION]
for span in exporter.get_finished_spans()
if span.attributes and DURATION in span.attributes
]
class _Clock:
"""A clock that advances by exactly ELAPSED_SECONDS on its second read."""
def __init__(self):
self.calls = 0
def __call__(self):
self.calls += 1
return 1000.0 if self.calls == 1 else 1000.0 + ELAPSED_SECONDS
def _wrapped(tracer, wrapped):
return async_mod.async_general_wrap(
"crawl",
"0.0.0-test",
"test-env",
"test-app",
tracer,
pricing_info={},
capture_message_content=False,
metrics={},
disable_metrics=True,
)(wrapped, None, (), {})
def test_normal_await_records_seconds():
tracer, exporter = _tracer_with_exporter()
async def _call(*_args, **_kwargs):
return {"success": True}
with patch.object(async_mod.time, "time", _Clock()):
asyncio.run(_wrapped(tracer, _call))
recorded = _recorded_durations(exporter)
assert recorded, "the async wrapper recorded no duration at all"
for value in recorded:
# 2500.0 before the fix: milliseconds, 1000x past the seconds ladder
# openlit declares for this histogram.
assert value == pytest.approx(ELAPSED_SECONDS), (
f"duration is declared in seconds, got {value!r}"
)
def test_error_path_records_seconds():
tracer, exporter = _tracer_with_exporter()
async def _raises(*_args, **_kwargs):
raise RuntimeError("boom")
with patch.object(async_mod.time, "time", _Clock()):
with pytest.raises(RuntimeError):
asyncio.run(_wrapped(tracer, _raises))
recorded = _recorded_durations(exporter)
assert recorded, "the error path recorded no duration at all"
for value in recorded:
assert value == pytest.approx(ELAPSED_SECONDS), (
f"the error path must use the same unit, got {value!r}"
)
def test_internal_millisecond_bookkeeping_is_not_the_attribute():
"""openlit keeps an internal millisecond aggregate; that is not the attribute.
A purely static guard cannot separate the two, because both derive from the
same elapsed value in the same function. This pins the distinction: the
semconv attribute is seconds while the internal list stays in milliseconds.
"""
tracer, exporter = _tracer_with_exporter()
metrics: dict = {}
async def _call(*_args, **_kwargs):
return {"success": True}
with patch.object(async_mod.time, "time", _Clock()):
asyncio.run(
async_mod.async_general_wrap(
"crawl",
"0.0.0-test",
"test-env",
"test-app",
tracer,
pricing_info={},
capture_message_content=False,
metrics=metrics,
disable_metrics=False,
)(_call, None, (), {})
)
for value in _recorded_durations(exporter):
assert value == pytest.approx(ELAPSED_SECONDS)
for key, samples in metrics.items():
if key.endswith("duration"):
for sample in samples:
assert sample == pytest.approx(ELAPSED_SECONDS * 1000), (
f"internal {key} is milliseconds by design, got {sample!r}"
)