A dashboard can be beautiful, green, and almost completely unhelpful. The test is not whether a system emits telemetry. The test is whether an engineer can move from a vague symptom to a defensible explanation while the incident is still happening.
Useful observability preserves context: which customer, request, version, region, dependency, and resource condition participated in the behavior.
Instrument decisions, not decoration
Begin with the questions operators need to answer. Is the problem global or isolated? Did it begin after a change? Is work failing, slowing, or never arriving? Which dependency is consuming the budget?
Metrics reveal patterns, traces expose paths, and logs preserve detail. Their value compounds when identifiers and semantic conventions connect them.
Signals need owners
Telemetry without stewardship decays. Assign ownership, review high-cardinality cost, remove alerts that do not change behavior, and test whether runbooks still match reality. The goal is a living operational model—not an expensive archive of everything the system once said.