production-monitoring

name	production-monitoring
description	Production observability — OpenTelemetry traces, structured logging, metrics, alerting, health endpoints, and SLO definition. Use this skill when the user mentions monitoring, observability, logging, metrics, traces, alerts, SLOs, or says /production monitoring. Triggers on observability discussions, OTEL setup, structured logging configuration, Prometheus/Grafana setup, or alerting rules.

Production Monitoring and Observability

This skill encodes battle-tested observability patterns for production services. Every recommendation comes from real incidents — the ones where you stared at a dashboard that showed nothing useful while users were screaming. Observability is not a feature you bolt on after launch. It is the foundation you build on from day one.

---

1. The Three Pillars of Observability

Observability is not "having logs." It is the ability to ask arbitrary questions about your system's behavior without deploying new code. The three pillars work together — none is sufficient alone.

Pillar	What It Tells You	Example
Logs	What happened — discrete events with context	"User X login failed: expired token"
Metrics	How the system behaves now — aggregated numbers over time	"p99 latency is 450ms and rising"
Traces	Why something is slow — a request's journey across services	"Postgres query in user-service took 2.3s"

How they connect: An alert fires on a metric (error rate > 1%). You filter logs by the time window to see what errors occurred. You grab a trace ID from the logs and follow the trace to the slow service. You fix it and verify the metric recovers. Without all three, you are flying blind.

---

2. Structured Logging

Unstructured logs (print("something went wrong")) are useless in production. You cannot filter, aggregate, or dashboard them.

Python: structlog Setup

import structlog, logging

def configure_logging(environment: str) -> None:
    processors: list[structlog.types.Processor] = [
        structlog.contextvars.merge_contextvars,
        structlog.processors.add_log_level,
        structlog.processors.StackInfoRenderer(),
        structlog.processors.TimeStamper(fmt="iso"),
        structlog.processors.format_exc_info,
    ]
    renderer = (structlog.processors.JSONRenderer() if environment == "production"
                else structlog.dev.ConsoleRenderer())

    structlog.configure(
        processors=[*processors, structlog.stdlib.ProcessorFormatter.wrap_for_formatter],
        logger_factory=structlog.stdlib.LoggerFactory(),
        wrapper_class=structlog.stdlib.BoundLogger,
        cache_logger_on_first_use=True,
    )
    formatter = structlog.stdlib.ProcessorFormatter(
        processors=[structlog.stdlib.ProcessorFormatter.remove_processors_meta, renderer],
    )
    handler = logging.StreamHandler()
    handler.setFormatter(formatter)
    root = logging.getLogger()
    root.handlers.clear()
    root.addHandler(handler)
    root.setLevel(logging.INFO)

    # Silence noisy libraries
    for lib in ("uvicorn.access", "httpx", "sqlalchemy.engine"):
        logging.getLogger(lib).setLevel(logging.WARNING)

Node.js: pino Setup

import pino from "pino";

const logger = pino({
  level: process.env.LOG_LEVEL || "info",
  transport: process.env.NODE_ENV !== "production"
    ? { target: "pino-pretty", options: { colorize: true } } : undefined,
  base: { service: process.env.SERVICE_NAME || "my-service" },
  redact: ["req.headers.authorization", "req.headers.cookie", "*.password", "*.token"],
});
export default logger;

Log Levels Discipline

Log levels are a contract with your on-call engineers, not a suggestion.

Level	Meaning	Alert?	Example
ERROR	Needs human attention. An alert should fire.	Yes	Database connection failed, payment processing error, unhandled exception
WARNING	Something unexpected happened but was handled.	No	Rate limit hit, cache miss fallback, deprecated API called
INFO	Business events. The happy path.	No	User created, order placed, deployment started
DEBUG	Developer diagnostics. Never in production.	No	SQL query text, request/response bodies, internal state

Rules:

• If nobody will read it, do not log it
• If it is ERROR, there must be a corresponding alert. Otherwise it is WARNING
• DEBUG logs in production are a performance tax with zero value — disable them

NEVER Log / ALWAYS Include

NEVER log: passwords, tokens/authorization headers, credit card numbers, SSNs/PII, raw request bodies (may contain secrets).

ALWAYS include in every log line:

logger.info("order_placed",
    request_id="req-abc123",    # Ties to HTTP request
    trace_id="trace-def456",    # Ties to distributed trace
    user_id="user-789",         # Who triggered this
    order_id="order-012",       # What business entity
    amount=99.99, currency="USD",
    service="order-service",    # Which service emitted this
)

Correlation IDs Across Services

Every request gets a unique ID at the edge. Pass it downstream in headers. Include it in every log line.

import uuid, structlog
from starlette.types import ASGIApp, Receive, Scope, Send

class CorrelationIDMiddleware:
    def __init__(self, app: ASGIApp) -> None:
        self.app = app

    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
        if scope["type"] not in ("http", "websocket"):
            await self.app(scope, receive, send)
            return
        headers = dict(scope.get("headers", []))
        request_id = (headers.get(b"x-request-id", b"").decode()
                      or headers.get(b"x-correlation-id", b"").decode()
                      or str(uuid.uuid4()))
        structlog.contextvars.clear_contextvars()
        structlog.contextvars.bind_contextvars(request_id=request_id)

        async def send_with_id(message):
            if message["type"] == "http.response.start":
                h = list(message.get("headers", []))
                h.append((b"x-request-id", request_id.encode()))
                message["headers"] = h
            await send(message)
        await self.app(scope, receive, send_with_id)

# Propagate to downstream services
async def call_downstream(client: httpx.AsyncClient, url: str):
    rid = structlog.contextvars.get_contextvars().get("request_id", "unknown")
    return await client.get(url, headers={"X-Request-ID": rid})

---

3. OpenTelemetry (OTEL) Setup

OpenTelemetry is the vendor-neutral standard. Instrument once, export to Jaeger, Tempo, Datadog, or any OTLP backend.

Python Dependencies

opentelemetry-api, opentelemetry-sdk, opentelemetry-exporter-otlp-proto-grpc
opentelemetry-instrumentation-fastapi, -sqlalchemy, -httpx, -redis

Complete FastAPI Integration

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.resources import Resource, SERVICE_NAME, SERVICE_VERSION
from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor
from opentelemetry.instrumentation.sqlalchemy import SQLAlchemyInstrumentor
from opentelemetry.instrumentation.httpx import HTTPXClientInstrumentor
from opentelemetry.instrumentation.redis import RedisInstrumentor

def configure_tracing(service_name: str, service_version: str, otlp_endpoint: str) -> None:
    resource = Resource.create({
        SERVICE_NAME: service_name, SERVICE_VERSION: service_version,
        "deployment.environment": os.getenv("ENVIRONMENT", "development"),
    })
    provider = TracerProvider(resource=resource)
    provider.add_span_processor(BatchSpanProcessor(
        OTLPSpanExporter(endpoint=otlp_endpoint, insecure=True)
    ))
    trace.set_tracer_provider(provider)
    # Auto-instrument: creates spans for every request, SQL query, HTTP call, Redis command
    SQLAlchemyInstrumentor().instrument()
    HTTPXClientInstrumentor().instrument()
    RedisInstrumentor().instrument()

@asynccontextmanager
async def lifespan(app: FastAPI):
    configure_tracing("order-service", settings.app_version, settings.otlp_endpoint)
    yield
    trace.get_tracer_provider().shutdown()

app = FastAPI(title="Order Service", lifespan=lifespan)
FastAPIInstrumentor.instrument_app(app)

Manual Spans for Business Logic

Auto-instrumentation covers libraries. The most valuable spans are on your business logic.

tracer = trace.get_tracer(__name__)

async def process_order(order_id: str, user_id: str) -> Order:
    with tracer.start_as_current_span("process_order",
        attributes={"order.id": order_id, "user.id": user_id}) as span:
        with tracer.start_as_current_span("validate_inventory"):
            if not await check_inventory(order_id):
                span.set_status(trace.StatusCode.ERROR, "Insufficient inventory")
                raise InsufficientInventoryError(order_id)
        with tracer.start_as_current_span("charge_payment") as ps:
            payment = await charge_payment(order_id)
            ps.set_attribute("payment.amount", payment.amount)
        span.add_event("order_completed", attributes={"order.total": payment.amount})
        return order

Context Propagation (W3C TraceContext)

Instrumented HTTP clients inject traceparent headers automatically. For non-instrumented clients:

from opentelemetry.propagate import inject
headers = {}
inject(headers)  # Adds traceparent + tracestate
response = await some_client.get(url, headers=headers)

Exporter Configuration

# Jaeger: docker run -d -p 16686:16686 -p 4317:4317 jaegertracing/all-in-one:latest
OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317   # Jaeger / Tempo / OTEL Collector

---

4. Metrics (RED Method)

The RED method gives you the three metrics that matter most for request-driven services. If you measure nothing else, measure these.

• Rate — requests per second (throughput)
• Errors — error rate as a percentage (4xx and 5xx)
• Duration — latency distribution (p50, p95, p99)

Prometheus Client Setup (Python)

from prometheus_client import Counter, Histogram, Gauge, generate_latest, CONTENT_TYPE_LATEST
from starlette.types import ASGIApp, Receive, Scope, Send
from starlette.responses import Response
import time, re

REQUEST_COUNT = Counter("http_requests_total", "Total HTTP requests",
                        ["method", "endpoint", "status_code"])
REQUEST_DURATION = Histogram("http_request_duration_seconds", "Request duration",
    ["method", "endpoint"],
    buckets=[0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1.0, 2.5, 5.0, 10.0])
REQUESTS_IN_PROGRESS = Gauge("http_requests_in_progress", "In-flight requests",
                             ["method", "endpoint"])

class PrometheusMiddleware:
    def __init__(self, app: ASGIApp) -> None:
        self.app = app

    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
        if scope["type"] != "http":
            await self.app(scope, receive, send)
            return
        method, path = scope["method"], scope["path"]
        # Normalize: /users/123 -> /users/{id} to prevent cardinality explosion
        endpoint = re.sub(r"/\d+", "/{id}", re.sub(
            r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}", "{id}", path))
        REQUESTS_IN_PROGRESS.labels(method=method, endpoint=endpoint).inc()
        start, status_code = time.perf_counter(), 500

        async def send_wrapper(message):
            nonlocal status_code
            if message["type"] == "http.response.start":
                status_code = message["status"]
            await send(message)
        try:
            await self.app(scope, receive, send_wrapper)
        finally:
            REQUEST_COUNT.labels(method=method, endpoint=endpoint, status_code=status_code).inc()
            REQUEST_DURATION.labels(method=method, endpoint=endpoint).observe(time.perf_counter() - start)
            REQUESTS_IN_PROGRESS.labels(method=method, endpoint=endpoint).dec()

# Scrape target for Prometheus
async def metrics_endpoint(request):
    return Response(content=generate_latest(), media_type=CONTENT_TYPE_LATEST)

Custom Business Metrics

Technical metrics tell you the system is healthy. Business metrics tell you the business is healthy.

ORDERS_PLACED = Counter("orders_placed_total", "Total orders", ["payment_method", "region"])
ORDER_VALUE = Histogram("order_value_dollars", "Order value",
                        buckets=[10, 25, 50, 100, 250, 500, 1000, 5000])
ACTIVE_USERS = Gauge("active_users_current", "Currently active users")

Cardinality Awareness

High cardinality kills Prometheus. This is the #1 Prometheus misconfiguration.

# DANGEROUS — user_id has millions of values = millions of time series = OOM
Counter("http_requests_total", "...", ["method", "endpoint", "user_id"])  # cardinality bomb

# SAFE — all labels have bounded values