Production Observability skill

Production Observability

Instrument user-visible operations and operational boundaries, not every method.

Baseline

• Add Actuator and one metrics registry selected by the deployment platform.
• Use Micrometer Observation for application metrics and traces.
• Use Micrometer Tracing with the chosen bridge; do not mix tracing APIs throughout business code.
• Export through OTLP when the platform standardizes on OpenTelemetry collectors.
• Expose only required actuator endpoints and secure every non-public endpoint.

Observation conventions

• Name observations by stable operation, such as orders.create.
• Keep metric tags low-cardinality: method, outcome, region, or bounded status.
• Put request IDs, user IDs, order IDs, and exception messages only in traces or logs.
• Propagate context across @Async, executor, and Reactor boundaries.
• Record latency, throughput, failures, and saturation for every external dependency.

Health and readiness

• Keep liveness independent from remote systems so a dependency outage does not restart every pod.
• Put required dependencies in readiness groups.
• Write custom health indicators only for dependencies that affect traffic acceptance.
• Set explicit timeouts on health checks.

Logging and alerts

• Emit structured logs with trace and span correlation.
• Redact credentials, tokens, personal data, prompts, and payloads by default.
• Alert on symptoms tied to service objectives, not raw metric noise.
• Include runbook links and enough dimensions to identify the affected service and dependency.

Examples

• See examples/good-observation.java and examples/bad-observation.java.

Gotchas

• Agent tags metrics with user or entity IDs - this creates unbounded cardinality.
• Agent exposes every actuator endpoint publicly - expose the minimum and secure it.
• Agent makes liveness depend on the database - dependency outages then cause restart loops.
• Agent logs request bodies and tokens for debugging - redact sensitive data before emission.
• Agent creates spans but loses context in async work - configure context propagation explicitly.