impulse-telemetry 0.2.0

Observability SDK for Impulse microservices

impulse_telemetry

Observability SDK for Impulse microservices. One init() call wires up distributed tracing, Prometheus metrics, structured logging, and FastAPI middleware.

Install

pip install impulse-telemetry

# With SQLAlchemy / Redis / Celery auto-instrumentation
pip install "impulse-telemetry[extras]"

# With ML data quality checks (pandas required)
pip install "impulse-telemetry[ml]"

Quickstart

from fastapi import FastAPI
from impulse_telemetry import init

app = FastAPI()
init(app, service="my-service", version="1.0.0", env="production")

Every HTTP request now emits RED metrics, a distributed trace, and a structured JSON log line — automatically.

init() parameters

Parameter	Default	Description
app	None	FastAPI app. Pass to enable HTTP middleware.
service	required	Service name — appears on all telemetry.
version	"0.0.0"	Semver string tagged on all telemetry.
env	$ENV	"production" | "staging" | "dev"
otlp_endpoint	$OTEL_EXPORTER_OTLP_ENDPOINT	OTLP Collector gRPC address.
prometheus_port	None	Expose /metrics on this port (for workers without HTTP).
log_level	"INFO"	"DEBUG" | "INFO" | "WARNING" | "ERROR"
enable_ml	False	Register ML-specific metric instruments.

Logging

from impulse_telemetry.logging import get_logger

log = get_logger(__name__)
log.info("prediction_served", model_id="rec-v3", latency_ms=43)
# {"service":"my-service","trace_id":"abc..","event":"prediction_served","model_id":"rec-v3",...}

trace_id and span_id are injected automatically from the active OTEL span.

Per-request context

Bind fields once per request — they appear on every subsequent log line for that request.

from impulse_telemetry.logging import bind_request_context, clear_request_context

token = bind_request_context(user_id="usr_123", request_id="req_456")
log.info("doing_work")   # user_id + request_id injected automatically
clear_request_context(token)

Metrics

Prometheus metrics are pre-registered and labeled with service and env.

Automatic (via middleware)

Metric	Type	Description
http_requests_total	Counter	Requests by route, method, status
http_request_errors_total	Counter	4xx / 5xx responses
http_request_duration_seconds	Histogram	Request latency
http_active_requests	Gauge	In-flight requests

Manual

from impulse_telemetry.metrics import get_metrics

m = get_metrics()
m.dependency_latency.labels(**m.labels(dependency="redis")).observe(elapsed)
m.dependency_errors.labels(**m.labels(dependency="redis")).inc()

Available instruments: rate_limiter_hits, rate_limiter_remaining, dependency_latency, dependency_errors, queue_depth, circuit_breaker.

Tracing

from impulse_telemetry.tracing import get_tracer, inject_headers

tracer = get_tracer(__name__)

with tracer.start_as_current_span("my_operation") as span:
    span.set_attribute("key", "value")
    headers = inject_headers({})   # propagate W3C traceparent to downstream
    requests.get("http://other-service/api", headers=headers)

Auto-instrumented libraries (when installed): requests, httpx, SQLAlchemy, Redis, Celery.

ML Monitoring

Enable at startup with enable_ml=True.

from impulse_telemetry.ml import MLMonitor

monitor = MLMonitor(model_id="rec-v3", user_id=user_id)

Inference tracing

# Context manager
with monitor.inference(features=df) as span:
    result = model.predict(df)
    span.record_output(result)

# Decorator
@monitor.trace
def predict(features):
    return model.predict(features)

Both record inference latency, error count, and run data quality checks (missing values, schema violations) automatically.

Drift & performance metrics

# Call from batch evaluation jobs
monitor.record_drift("age", score=0.18, metric="psi")
monitor.record_performance(rmse=0.042, precision=0.87, recall=0.81)

ML metric instruments

Metric	Type	Description
ml_inference_duration_seconds	Histogram	Per-model inference latency
ml_inference_requests_total	Counter	Inference count
ml_inference_errors_total	Counter	Inference errors
ml_missing_feature_rate	Gauge	Missing values per column
ml_schema_violations_total	Counter	Schema violation count
ml_feature_drift	Gauge	Feature drift score
ml_prediction_drift	Gauge	Prediction drift score
ml_rmse	Gauge	Rolling RMSE
ml_precision	Gauge	Rolling precision
ml_recall	Gauge	Rolling recall

Background workers

For services without an HTTP server, expose metrics on a dedicated port:

init(service="ingest-worker", prometheus_port=9090)
# Prometheus scrapes localhost:9090/metrics

Environment variables

Variable	Description
OTEL_EXPORTER_OTLP_ENDPOINT	OTLP Collector gRPC address
ENV	Deployment environment (production, staging, dev)

See also

examples.py — minimal, runnable code for every feature.