observability

Add RED signals (requests, errors, duration) to every user-reachable path, traces across inbound and outbound boundaries, and dependency health metrics. Use structured logs with stable event names, severity, outcome, and trace correlation IDs.

Structured logging uses a stable schema with typed fields and semantic event names like checkout.payment_failed rather than prose messages. OpenTelemetry semantic conventions standardize service, HTTP, database, and exception attributes across your telemetry.

High-cardinality values like user IDs, request IDs, and paths explode metric cardinality, which is a real production cost in time-series systems. Put those values in logs or traces instead, and keep metric label sets bounded.

No. Liveness probes must never depend on external systems, or a dependency outage will cause orchestrators to restart healthy containers. Put dependency checks in the readiness probe instead.

Alerts should fire on user impact, SLO burn rate, or actionable dependency failure, each linking to a runbook with action and escalation steps. Tune noisy alerts by signal, owner, or threshold rather than silencing them.

A prototype may defer observability only if the deferral is recorded and the path is instrumented before real users reach it. Local-only scripts and libraries with no operational surface are also out of scope.