Chapters 3 and 4 could say what the hosts were doing on a Saturday morning; nothing could say what checkout was doing. This chapter moves measurement inside the code — prometheus_client in the bookings service, the four golden signals at its boundaries, label discipline so the new power doesn't become an outage, exemplars to link a p99 spike to one real request — and exporters, probes, and log mining for everything Harborline runs but didn't write.
5 topics
Mara pairs with the bookings team for a week, and the deliverable is a metric: harborline_checkout_duration_seconds, the first number that can answer the Saturday complaint at the service level rather than the host level. Getting there takes three instrument types, one mounted endpoint, and a live demo of the gunicorn multiprocess trap — the deployment mistake everyone makes exactly once.
Five topics carry the chapter: the client library's mechanics and its multiprocess price list; the four golden signals that decide what deserves a metric at all; label discipline, where one multiplication separates a 7,000-series histogram from a 1.2-million-series outage; exemplars, the bridge from an aggregate spike back to one traceable request; and the decision ladder — exporter, probe, log parser — for Postgres, Redis, nginx, and everything else Harborline cannot edit.
Where instrumentation lives — the app keeps score, Prometheus pulls it