Capacity Planning Fundamentals

Capacity planning ensures your infrastructure has enough resources to meet current and future demand without over-provisioning. Done well, it balances cost against risk — spending enough to handle traffic spikes without wasting money on idle capacity.

Why Capacity Planning Matters

Without capacity planning, you're always reacting. Traffic grows, things slow down, customers get frustrated, and you scramble to add resources. With capacity planning, you forecast growth, plan scaling in advance, and make data-driven decisions about infrastructure investment.

The cost of under-provisioning is downtime, slow performance, and lost revenue. The cost of over-provisioning is wasted cloud spend — often 20-40% higher than necessary.

Key Concepts

Headroom — The spare capacity available above current utilization. A service with 60% CPU utilization has 40% headroom. Critical services need more headroom than batch jobs.

Utilization — What percentage of total capacity is being used. Track P50, P90, and P99 utilization to understand typical load, peak load, and extreme spikes.

Capacity — The total resources available: CPU cores, memory, disk IOPS, network bandwidth. Capacity is not static — you can scale up (bigger instances) or scale out (more instances).

Demand — The load on your system: requests per second, active users, data volume. Demand drives utilization, which determines whether capacity is sufficient.

The Planning Horizon

Capacity planning operates on three time scales:

Short-term (days to weeks) — Respond to immediate trends. Add instances, adjust auto-scaling thresholds, optimize expensive queries.

Medium-term (months) — Anticipate planned growth from product launches, marketing campaigns, or seasonal patterns. Provision additional capacity.

Long-term (quarters to years) — Strategic infrastructure investments. Cloud provider commitments, data center capacity, architecture changes.

Load Forecasting Methods

Forecasting future load is the core technical challenge of capacity planning. The right forecasting method depends on your data's characteristics: trend, seasonality, and noise.

Linear Regression

Linear regression fits a straight line through historical data points and extends it into the future. It's the simplest forecasting method and works well when your traffic grows at a consistent rate.

Best for: Steady, predictable growth (e.g., a mature product with stable month-over-month increases).

Limitations: Doesn't handle seasonality or sudden trend changes. A spike on Black Friday will pull the regression line up and inflate forecasts for the rest of the year.

Seasonal Decomposition

Seasonal decomposition separates a time series into three components: trend (long-term direction), seasonal (recurring patterns), and residual (random noise). The forecast combines projected trend with the historical seasonal pattern.

Best for: Traffic with clear daily, weekly, or annual patterns (e.g., a B2B SaaS with weekday peaks and weekend troughs).

Implementation: Identify the seasonality period (7 days for weekly, 24 hours for daily), apply decomposition to extract each component, then forecast each component separately.

Double Exponential Smoothing

This method uses two smoothing parameters: one for the level (current baseline) and one for the trend (direction of change). It adapts to changing trends by giving more weight to recent observations.

Best for: When trends are shifting — a product in rapid growth, or a declining legacy service.

Limitations: Still doesn't handle seasonality natively. For seasonal data, use triple exponential smoothing (Holt-Winters).

Auto-Method Selection

The toolkit's auto-forecast feature tests all available methods against your data and selects the one with the best fit. It evaluates each method using R-squared and mean absolute error, then returns the best-performing forecast along with confidence intervals that widen with distance.

Confidence Intervals

Every forecast is uncertain, and that uncertainty grows the further you project. Always report forecasts with confidence intervals (typically 80% and 95%). Planning for the upper bound of the 95% confidence interval provides a safety margin.