Observability

Observability is the ability to understand what a system is doing from the outside by examining its outputs - without needing to modify the system or guess at internals. The three pillars are logs (what happened), metrics (how the system is performing), and traces (where time is spent across service boundaries). These pillars are only useful when correlated - a spike in your p99 metric should link to traces, and those traces should link to logs. Invest in correlation from day one, not as a retrofit.

---

When to use this skill

Trigger this skill when the user:

• Adds structured logging to a service (pino, winston, log4j, Python logging)
• Instruments code with OpenTelemetry or a vendor SDK (Datadog, New Relic, Honeycomb)
• Defines SLIs, SLOs, or error budgets for a service
• Builds a Grafana or Datadog dashboard
• Writes Prometheus alerting rules or configures PagerDuty/Opsgenie routing
• Implements distributed tracing (spans, context propagation, sampling)
• Responds to alert fatigue or on-call burnout
• Tracks an incident and needs to correlate logs/traces/metrics

Do NOT trigger this skill for:

• Pure infrastructure provisioning (Terraform, Kubernetes YAML) - those are ops/IaC concerns
• Application performance profiling of CPU/memory at the code level (use a performance-engineering skill)

---

Key principles

1. Structured logging always - Every log line should be machine-parseable JSON with consistent fields. Plain-text logs cannot be queried, filtered, or aggregated at scale. Correlation IDs are non-negotiable.

2. USE for resources, RED for services - Resources (CPU, memory, connections) are measured with Utilization/Saturation/Errors. Services (APIs, queues) are measured with Rate/Errors/Duration. Knowing which method applies tells you which metrics to instrument before you write a single line of code.

3. Instrument at boundaries - Service ingress/egress, database calls, external HTTP calls, and message queue produce/consume operations are the minimum instrumentation surface. Everything else is optional until proven necessary.

4. Alert on symptoms, not causes - Alert when users are impacted (high error rate, high latency). Do not page on CPU at 80% or a memory warning - those are causes to investigate, not symptoms to wake someone up for.

5. SLOs drive decisions - Every reliability trade-off should reference an error budget. If budget is healthy, ship features. If budget is burning, stop and fix reliability. SLOs without error budgets are just numbers on a slide.

---

Core concepts

The three pillars

Pillar	Question answered	What it gives you
Logs	What happened?	Detailed event records, debug context, audit trails
Metrics	How is the system performing?	Aggregated numbers over time, dashboards, alerting
Traces	Where did time go?	Request flow across services, latency attribution

Cardinality

Every unique combination of label values in a metric creates a new time series in your metrics backend. user_id as a metric label will create millions of time series and kill Prometheus. Keep metric label cardinality under ~100 unique values per label. Use logs or traces for high-cardinality data (user IDs, request IDs, emails).

Exemplars

Exemplars are trace IDs embedded in metric data points. When you see a p99 spike on a histogram, an exemplar lets you jump directly to a trace that caused it. OpenTelemetry and Prometheus support exemplars natively. Enable them - they are the bridge between metrics and traces.

Context propagation

Context propagation is the mechanism by which a trace ID flows through service boundaries. The W3C traceparent header is the standard format. Every service must: extract the header on ingress, attach it to async context, and inject it into all outbound calls. Failing to propagate breaks trace continuity silently.

SLI / SLO / Error budget

• SLI (Service Level Indicator): A measurement of service behavior users care about. Example: successful_requests / total_requests
• SLO (Service Level Objective): A target for an SLI over a time window. Example: 99.9% of requests succeed within 300ms, measured over 30 days
• Error budget: 1 - SLO. For a 99.9% SLO, the budget is 0.1% - about 43 minutes of downtime per month. Burn rate measures how fast you consume it.

---

Common tasks

Set up structured logging

Use pino for Node.js (fastest), winston for flexibility. Always include a correlation ID middleware that attaches traceId to every log automatically.

// logger.ts - pino with correlation ID support
import pino from 'pino';

export const logger = pino({
  level: process.env.LOG_LEVEL ?? 'info',
  base: {
    service: process.env.SERVICE_NAME ?? 'unknown',
    version: process.env.SERVICE_VERSION ?? '0.0.0',
  },
  timestamp: pino.stdTimeFunctions.isoTime,
  redact: ['req.headers.authorization', 'body.password', 'body.token'],
});

// Express middleware - binds traceId to every child logger in the request scope
export function loggerMiddleware(req: Request, res: Response, next: NextFunction) {
  const traceId = req.headers['traceparent'] as string
    ?? req.headers['x-request-id'] as string
    ?? crypto.randomUUID();

  req.log = logger.child({ traceId, method: req.method, path: req.path });
  res.setHeader('x-request-id', traceId);
  next();
}

// Usage in a route handler
app.post('/orders', async (req, res) => {
  req.log.info({ orderId: body.id }, 'Processing order');
  try {
    const result = await orderService.create(body);
    req.log.info({ orderId: result.id, durationMs: Date.now() - start }, 'Order created');
    res.json(result);
  } catch (err) {
    req.log.error({ err, orderId: body.id }, 'Order creation failed');
    res.status(500).json({ error: 'internal_error' });
  }
});

Instrument with OpenTelemetry

Use the Node.js SDK with auto-instrumentation for HTTP, Express, and common DB clients. Add manual spans only for business-critical operations.

// instrumentation.ts - must be loaded before any other module (Node --require flag)
import { NodeSDK } from '@opentelemetry/sdk-node';
import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';
import { OTLPMetricExporter } from '@opentelemetry/exporter-metrics-otlp-http';
import { PeriodicExportingMetricReader } from '@opentelemetry/sdk-metrics';
import { getNodeAutoInstrumentations } from '@opentelemetry/auto-instrumentations-node';
import { ParentBasedSampler, TraceIdRatioBased } from '@opentelemetry/sdk-trace-node';

const sdk = new NodeSDK({
  serviceName: process.env.SERVICE_NAME ?? 'my-service',
  traceExporter: new OTLPTraceExporter({
    url: process.env.OTEL_EXPORTER_OTLP_ENDPOINT ?? 'http://localhost:4318/v1/traces',
  }),
  metricReader: new PeriodicExportingMetricReader({
    exporter: new OTLPMetricExporter(),
    exportIntervalMillis: 15_000,
  }),
  sampler: new ParentBasedSampler({
    root: new TraceIdRatioBased(0.1), // 10% head-based sampling
  }),
  instrumentations: [getNodeAutoInstrumentations()],
});

sdk.start();
process.on('SIGTERM', () => sdk.shutdown());

// Manual span for a business operation
import { trace, SpanStatusCode } from '@opentelemetry/api';

const tracer = trace.getTracer('order-service');

async function processPayment(orderId: string, amount: number) {
  return tracer.startActiveSpan('payment.process', async (span) => {
    span.setAttributes({ 'order.id': orderId, 'payment.amount': amount });
    try {
      const result = await stripe.charges.create({ amount, currency: 'usd' });
      span.setStatus({ code: SpanStatusCode.OK });
      return result;
    } catch (err) {
      span.setStatus({ code: SpanStatusCode.ERROR, message: (err as Error).message });
      span.recordException(err as Error);
      throw err;
    } finally {
      span.end();
    }
  });
}

Load instrumentation.ts before your app with node --require ./dist/instrumentation.js server.js. See references/opentelemetry-setup.md for exporters, processors, and Python setup.

Define SLIs and SLOs

Define SLIs from the user's perspective first, then map to metrics you can measure.

# slos.yaml - document alongside your service
service: order-api
slos:
  # Availability: are requests succeeding?
  - name: availability
    description: Fraction of requests that return non-5xx responses
    sli: successful_requests / total_requests  # status < 500
    target: 99.9%
    window: 30d
    error_budget_minutes: 43.8

  # Latency: are requests fast enough?
  - name: latency-p99
    description: 99th percentile of request duration under 500ms
    sli: requests_under_500ms / total_requests
    target: 99.0%
    window: 30d

  # Correctness: are responses valid? (measured via synthetic probes or sampling)
  - name: correctness
    description: Fraction of order confirmations that pass integrity check
    sli: valid_order_confirmations / total_order_confirmations
    target: 99.95%
    window: 30d

SLO burn rate formulas:

error_budget       = 1 - slo_target        # 0.001 for 99.9%
burn_rate          = observed_error_rate / error_budget
time_to_exhaustion = window_hours / burn_rate

# Fast burn (page now): 14.4x - exhausts 30d budget in 2 days
# Slow burn (ticket):    3x   - exhausts 30d budget in 10 days

Create effective dashboards

Use the RED method layout. Eight to twelve panels per dashboard. Link to detail dashboards for drill-down rather than putting everything on one page.

Dashboard layout - <ServiceName> Overview
Row 1: [SLO Status: availability]  [Error Budget: X% remaining]  [Latency p99 SLO]
Row 2: [Request Rate (rps)]  [Error Rate (%)]  [Latency p50 / p95 / p99]
Row 3: [Errors by type/endpoint]  [Top slow endpoints]  [Upstream dependency latency]
Row 4: [CPU / Memory]  [DB connection pool]  [Queue depth / lag]

Grafana panel guidelines:

• Latency: use histogram_quantile, show p50/p95/p99 on same panel
• Error rate: rate(errors_total[5m]) / rate(requests_total[5m])
• Add deploy annotations (vertical lines) so you can correlate deployments with incidents
• Set panel thresholds to match your SLO targets (green/yellow/red)

Set up alerting without alert fatigue

Define severity tiers before writing a single rule. Map each tier to a routing target.

# Example Prometheus alerting rules (alerts.yaml)
groups:
  - name: order-api.slo
    rules:
      # P1: fast burn - exhausts 30d budget in 2 days
      - alert: HighErrorBudgetBurn
        expr: |
          (
            rate(http_requests_errors_total[1h]) /
            rate(http_requests_total[1h])
          ) > (14.4 * 0.001)
        for: 2m
        labels:
          severity: p1
          team: platform
        annotations:
          summary: "Error budget burning at 14x+ rate"
          runbook: "https://runbooks.internal/order-api/high-error-burn"
          dashboard: "https://grafana.internal/d/order-api"

      # P3: slow burn - ticket, investigate during business hours
      - alert: SlowErrorBudgetBurn
        expr: |
          (
            rate(http_requests_errors_total[6h]) /
            rate(http_requests_total[6h])
          ) > (3 * 0.001)
        for: 1h
        labels:
          severity: p3
          team: platform
        annotations:
          summary: "Error budget burning at 3x rate - investigate during business hours"

Routing rules (Opsgenie / PagerDuty):
  severity=p1 -> Page primary on-call immediately
  severity=p2 -> Page primary on-call during business hours, silent at night
  severity=p3 -> Create Jira ticket, no page
  severity=p4 -> Slack notification only

Every alert must have: a runbook link, an owner team, and a dashboard link. If an alert fires and nobody knows what to do, the runbook is missing.

Implement distributed tracing

Instrument at service boundaries. Propagate context via W3C traceparent. Add attributes that make traces searchable (user ID, order ID, tenant ID - as trace attributes, not metric labels).

// Propagate context in outbound HTTP calls (fetch wrapper)
import { context, propagation } from '@opentelemetry/api';

async function tracedFetch(url: string, options: RequestInit = {}): Promise<Response> {
  const headers: Record<string, string> = {
    ...(options.headers as Record<string, string>),
  };
  // Inject W3C traceparent + tracestate headers
  propagation.inject(context.active(), headers);
  return fetch(url, { ...options, headers });
}

// Propagate context from inbound messages (e.g. SQS / Kafka)
import { propagation, ROOT_CONTEXT } from '@opentelemetry/api';

function processMessage(message: QueueMessage) {
  // Extract trace context from message attributes
  const parentContext = propagation.extract(ROOT_CONTEXT, message.attributes ?? {});
  return context.with(parentContext, () => {
    return tracer.startActiveSpan('queue.process', (span) => {
      span.setAttributes({ 'messaging.message_id': message.id });
      // ... process message
      span.end();
    });
  });
}

Span attribute conventions (OpenTelemetry semantic conventions):

• HTTP: http.method, http.status_code, http.route, net.peer.name
• DB: db.system, db.name, db.operation, db.statement (sanitized)
• Business: order.id, user.id, payment.method (custom namespace)

Monitor error budgets and act on burn rates

Track burn rate over multiple windows to distinguish spikes from trends.

// Burn rate queries (Prometheus / Grafana)

// 1-hour burn rate (catches fast incidents)
const fastBurnRate = `
  (
    sum(rate(http_requests_errors_total[1h])) /
    sum(rate(http_requests_total[1h]))
  ) / 0.001
`;

// 6-hour burn rate (catches slow degradations)
const slowBurnRate = `
  (
    sum(rate(http_requests_errors_total[6h])) /
    sum(rate(http_requests_total[6h]))
  ) / 0.001
`;

// Remaining error budget (30-day rolling)
const budgetRemaining = `
  1 - (
    sum(increase(http_requests_errors_total[30d])) /
    sum(increase(http_requests_total[30d]))
  ) / 0.001
`;

Act on burn rates:

Burn rate	Action
> 14.4x (1h window)	Page immediately, declare incident
> 6x (6h window)	Page during business hours
> 3x (24h window)	Create reliability ticket, add to next sprint
< 1x	Budget healthy, normal feature development
Budget < 10% remaining	Freeze non-critical deploys, focus on reliability

---

Anti-patterns / common mistakes

Mistake	Why it's wrong	What to do instead
Logging unstructured plain text	Cannot be searched or aggregated at scale	Emit JSON with consistent fields and correlation ID
High-cardinality metric labels (user_id, request_id)	Creates millions of time series, kills Prometheus	Keep cardinality < 100 per label; use traces for high-cardinality data
Alerting on causes (CPU > 80%)	Wakes humans for non-user-impacting events	Alert on symptoms (error rate, latency SLO burn)
No sampling strategy for traces	100% trace collection at scale is cost-prohibitive	Start at 10% head-based, add tail-based for errors
SLOs without error budgets	SLO becomes a vanity target with no operational consequence	Define budget, burn rate thresholds, and what changes at each level
Missing runbooks on alerts	On-call doesn't know what to do, wasted time in incidents	Every alert ships with a runbook before it goes to production

---

Gotchas

1. Cardinality explosion kills Prometheus - Adding a label with high cardinality (user_id, request_id, IP address) creates a new time series per unique value. A single bad label can OOM a Prometheus instance overnight. Always check cardinality before adding labels; use traces or logs for high-cardinality data.

2. Context propagation breaks at async boundaries - In Node.js, if you use setTimeout, setImmediate, or create a new Promise chain without explicitly passing context.active(), the trace context is lost and spans appear as orphan roots. Use AsyncLocalStorage-aware frameworks or manually propagate context with context.with(ctx, fn).

3. 100% trace sampling in production is unsustainable - At any real scale, sampling every trace destroys budget and storage. Start at 10% head-based sampling with tail-based sampling for errors. The default AlwaysOnSampler in OTel SDKs is NOT suitable for production.

4. SLO burn rate alerts on short windows produce noise - A single spike in errors can trigger a "fast burn" alert that resolves in minutes. Pair fast-window alerts (1h) with slow-window alerts (6h) using multi-window alerting. Alert only when both windows exceed the threshold simultaneously.

5. Structured logging without redaction leaks secrets - pino and winston log entire objects by default. Passing req or body without a redact config will log Authorization headers, passwords, and tokens in plain text. Always configure the redact option before shipping to production.

---

References

• references/opentelemetry-setup.md - OTel SDK setup for Node.js and Python, exporters, processors, and sampling configuration

Load the references file when the task involves wiring up OpenTelemetry from scratch, configuring exporters, or setting up the collector pipeline. The skill above is enough for instrumentation patterns and SLO definitions.