Instrumentation

Instrumentation is the emitting side of observability: the code that makes your app talk about itself. No instrumentation, no data, and no dashboard can conjure it after the fact. It comes in two flavors. Auto-instrumentation patches known libraries at startup (the HTTP server, fetch, the Postgres driver, Redis) and hands you a usable trace and a request-rate chart for roughly zero lines of your own code. Manual instrumentation is the part only you can write: a span around 'generate invoice PDF', a counter for 'coupon rejected', an attribute for plan tier.

The order that works: switch on auto-instrumentation first for cheap coverage, then add manual spans and attributes at the edges of your own business logic, one incident at a time. The best instrumentation gets written the day after an outage, when you remember exactly which field you wished you had. Prefer fewer, fatter spans carrying lots of context over a swarm of thin ones that say nothing.

Gotchas: auto-instrumentation usually has to load before everything else (a preload flag, or a require at the very top of the entrypoint), and quietly does nothing if a single import beats it. It also costs real CPU and real money at traffic, so sample. And instrumentation is the main way PII escapes into a vendor: request bodies, auth headers, and full URLs with query strings are captured by default in some agents, so redact at the collector rather than trusting each service.