Capacity planning
Forecasting resource needs to provision enough capacity without overpaying. Balances headroom for spikes against idle cost. Driven by growth trends, seasonality, and SLOs. The complement to reactive autoscaling.
- Any growing service with SLOs.
- Budgeting and procurement.
- Trivial services with no growth.
- When pure serverless handles it.
Definition
Capacity planning estimates future demand (from trends and seasonality) and provisions resources to meet SLOs with a safety margin. It sets baselines, headroom, and triggers for autoscaling; without it, you either overpay or fail under load.
Why it exists
Autoscaling reacts; you still need a baseline and headroom, and to budget. Planning ensures you can handle growth and known events (Black Friday) without last-minute scrambling or overprovisioning.
How it works
Track utilization trends and peak patterns. Model demand (linear growth, seasonality). Provision a baseline + headroom (e.g. 30%). Reserve capacity for known spikes. Use autoscaling for the variable part. Revisit against actuals.
Scaling limits
Forecasts are wrong; build in margin and review. Long procurement lead times (reserved hardware) force early planning. Overplanning wastes money; underplanning fails SLOs.
Cost shape
The discipline that prevents both overpaying and outages. Blow-ups: ignoring seasonality and over-reserving.
Alternatives
- Pure autoscaling — Reactive but no baseline/budget control.
- Overprovision and forget — Safe but wasteful.
Failure modes
- Underestimating peak growth.
- Ignoring seasonality.
- No headroom for spikes.