notes on infra you can hear

After enough time on call, a deployment starts to have a sound. Not literally — though I’ve worked in datacenters where it did — but a recognizable rhythm on a dashboard. You glance at it the way a mechanic glances at an idling engine, and you know whether it’s right. You can’t always articulate what’s wrong, but you can tell that something is.

For me it composes from three signals stacked together. Request rate gives you the load, the baseline hum. P99 latency gives you the timbre — whether the system is straining or coasting. Error budget burn gives you the trend, the difference between a transient cough and a developing problem. Any one of them in isolation lies; together they harmonize, and you learn the chord progression of a healthy service.

The corollary is that the most valuable on-call shift is a boring one. Boredom at a dashboard isn’t wasted time — it’s the calibration that lets you notice, immediately, when the sound changes.