An overloaded dashboard warns of nothing. Here are the signals we watch on every production project.
A dashboard with forty charts gives a feeling of control and rarely warns of an incident. We keep five, and we alert on them.
Application error rate first: the share of failing requests, measured per route. A spike on a single route is almost always a regression from the last release.
Then latency at the 95th percentile, never the average. The average hides exactly the users who suffer most.
Resource saturation, backup freshness and queue processing delay complete the set. Five signals, five alert thresholds, and an on-call engineer who knows what to do with each.