Observability That Actually Saves Rounds: Logs, Traces, SLOs

99.99% uptime is 52 minutes of downtime a year — you don’t get there with vibes. One SLO per surface (p95 latency, error budget, checkout success), traces from edge to DB, and logs you can actually query at 3am.

Our stack: Sentry for exceptions with release health, structured JSON logs with request IDs, and a weekly error-budget review. Pager only fires on burn-rate alerts — everything else is a ticket, not a siren.

Start small: instrument the money path first (signup → pay → deliver), add synthetic checks from two regions, and write the runbook before the incident. Calm on-call is a feature you ship.