Capacity planning with percentile baselines

Engineering

Average CPU or bandwidth looks calm right up until the moment it is not. A host that idles at 20% for most of the day and hits 95% for ten minutes every hour is not a 20% host.

Use percentiles

We compute the 95th percentile of each metric over a rolling week and compare it to the provisioned limit. When p95 crosses 70%, the capacity review is triggered.

Watch the trend, not the point

A single high reading is noise. A p95 that climbs by two points every week is a plan for next quarter.

Keep headroom for failure

If one node in a three-node group fails, the other two absorb its load. Size so that p95 stays under 80% in that scenario.