Measure recovery, not only failure
Reliability metrics that ignore time-to-restore leave operators flying blind after the first alert.
Failure counts tell you something broke. Recovery metrics tell you whether the system — and the team — can return to a known-good state. Application analytics should emit restore markers: when traffic healed, when backlog drained, when the degraded feature flipped back.
Without those markers, postmortems invent timelines from chat. With them, you can compare incidents on mean time to restore and spot processes that routinely stall.
Instrument restore events with the same care you give error events. Pair them in reviews. Reliability improves when recovery is a first-class signal, not an afterthought in a spreadsheet.