Observe

Instrument so that the next failure is detected by a signal, not a user, and can be explained from the data without a redeploy. Every signal needs a question it answers.

Phase 1: Define what matters

• Name the user-facing outcomes and the questions an on-call engineer will ask: is it working, for whom is it failing, since when, and where in the path.
• Pick a few service-level indicators from those outcomes (success rate, latency percentiles, freshness, correctness), with targets when the team uses them.
• Read the existing telemetry stack, libraries, naming conventions and dashboards. Extend them; do not add a second logging or tracing library.

Phase 2: Instrument the path