Verify exact production traces and record rollout

This commit is contained in:
2026-08-11 11:29:59 +07:00
parent 6b8f7584ed
commit 59e6ad2d0d
46 changed files with 2795 additions and 290 deletions
+122
View File
@@ -0,0 +1,122 @@
from adapters.prometheus import PrometheusMetrics
from config import Settings
from rag.instrumentation import InstrumentedGenerator
from rag.metrics import (
CLARIFY_ASKED,
PROVIDER_FAILURE,
REQUEST_DURATION,
REQUESTS,
STAGE_DURATION,
TRACE_WRITE_FAILED,
InMemoryMetrics,
)
from rag.telemetry import (
configure_telemetry,
correlation_context,
request_span,
stage,
)
def test_prometheus_registers_domain_failures_histograms_and_bounds_labels():
metrics = PrometheusMetrics()
metrics.increment(TRACE_WRITE_FAILED)
metrics.increment(CLARIFY_ASKED, reason="raw-user-controlled-reason")
metrics.increment(
REQUESTS, method="POST", route="/v1/rag/query", status="2xx"
)
metrics.observe(
REQUEST_DURATION,
1.25,
method="POST",
route="/v1/rag/query",
status="2xx",
)
metrics.observe(STAGE_DURATION, 0.2, stage="retrieval", outcome="ok")
body, content_type = metrics.render()
rendered = body.decode("utf-8")
assert "application/openmetrics-text" in content_type
assert "duocthu_trace_write_failed_total 1.0" in rendered
assert 'duocthu_clarify_asked_total{reason="other"} 1.0' in rendered
assert "duocthu_request_duration_seconds_bucket" in rendered
assert "duocthu_stage_duration_seconds_bucket" in rendered
def test_stage_records_duration_without_requiring_an_otel_collector():
metrics = InMemoryMetrics()
configure_telemetry(Settings(otel_enabled=False), metrics)
with stage("grounding"):
pass
values = metrics.observed(STAGE_DURATION, stage="grounding", outcome="ok")
assert len(values) == 1
assert values[0] >= 0
class _FailingGenerator:
model_id = "fake"
def generate(self, system: str, user: str, schema: dict) -> str:
raise ConnectionError("provider offline")
def test_provider_failure_uses_bounded_classification_not_exception_text():
metrics = InMemoryMetrics()
generator = InstrumentedGenerator(_FailingGenerator(), metrics)
try:
generator.generate("system", "user", {})
except ConnectionError:
pass
else: # pragma: no cover - assertion guard
raise AssertionError("expected provider failure")
assert metrics.total(
PROVIDER_FAILURE,
provider="other",
operation="generate",
reason="provider_unavailable",
) == 1
def test_otel_server_and_stage_spans_share_trace_and_emit_metric_exemplar():
otel_trace = pytest.importorskip("opentelemetry.trace")
sdk_trace = pytest.importorskip("opentelemetry.sdk.trace")
exporter_module = pytest.importorskip(
"opentelemetry.sdk.trace.export.in_memory_span_exporter"
)
export_module = pytest.importorskip("opentelemetry.sdk.trace.export")
exporter = exporter_module.InMemorySpanExporter()
provider = sdk_trace.TracerProvider()
provider.add_span_processor(export_module.SimpleSpanProcessor(exporter))
otel_trace.set_tracer_provider(provider)
metrics = PrometheusMetrics()
configure_telemetry(
Settings(
otel_enabled=True,
otel_exporter_otlp_endpoint="http://127.0.0.1:4318/v1/traces",
),
metrics,
)
with correlation_context("req-span-test"):
with request_span("POST", "/v1/rag/query", {}):
with stage("retrieval"):
pass
provider.force_flush()
spans = exporter.get_finished_spans()
assert {item.name for item in spans} == {
"POST /v1/rag/query",
"rag.stage.retrieval",
}
assert len({item.context.trace_id for item in spans}) == 1
assert spans[-1].attributes["duocthu.correlation_id"] == "req-span-test"
rendered = metrics.render()[0].decode("utf-8")
assert "duocthu_stage_duration_seconds_bucket" in rendered
assert 'trace_id="' in rendered
import pytest