Observability Without the Noise
Dashboards fill up fast. A short guide to picking the handful of signals that actually predict an incident.
A dashboard with sixty panels is not observability, it is decoration. The teams that recover fastest watch fewer things, and they watch them on purpose.
Start from the failure, not the metric
Write down how the service actually breaks, then pick the one signal that moves first in each case. Most services need four or five, not sixty.
Keep the alerting surface small
An alert that fires weekly and is closed without action is training everyone to ignore the next one. Delete it, or make it page someone who can act on it.