sethdford Capacity Planning
Forecast resource needs and plan infrastructure to meet demand while optimizing costs.
Context
You are planning infrastructure capacity. Model growth, forecast peak load, size infrastructure for headroom. Read historical metrics, growth trends, and business projections.
Domain Context
Based on capacity planning and performance modeling:
- Utilization Target: 70-80% typical for servers (avoid 100% - no headroom). Database: 70-80%. Network: 60-70% to avoid congestion.
- Headroom: Plan for 2-3x growth before re-architecture. At 70% utilization with 10% monthly growth, capacity lasts ~6 months.
- Seasonality: Traffic varies by season (Black Friday 10x normal). Plan for peaks; over-provision not necessary for short bursts if auto-scaling handles it.
- Growth Curve: Linear (startup phase), exponential (hypergrowth), plateau (mature). Different forecasting strategies per phase.
Instructions
Establish Baseline: Measure current usage (CPU, memory, storage, network bandwidth). Peak vs average. Growth rate over past 6-12 months.
Project Growth: Based on growth rate and business projections. Linear extrapolation for stable growth. Exponential model for hypergrowth startups. Add margin for uncertainty.
Calculate Required Capacity: For peak load + 20% headroom, how many servers needed? At 70% CPU utilization per server, peak_load / 0.7 = servers needed.
Model Scenarios: Best case (fast growth, more servers). Worst case (slow growth, fewer servers). Plan investment for mid-case; be ready to scale up/down.
Plan Upgrades: When do you need additional capacity? In 6 months? Plan lead time for procurement, setup, migration. Use auto-scaling to smooth spikes; only add permanent capacity when needed.
Anti-Patterns
- Over-Provisioning: Buy capacity for 5 years of growth. Result: underutilized, wasted cost. Guard: Plan for 1-2 years; cloud enables rapid scaling.
- No Growth Plan: Assume current load is constant. Result: surprised by capacity limits. Guard: Monitor trends; forecast quarterly.
- Ignoring Auto-Scaling: Manual provisioning for every 10% traffic increase. Result: operational burden, missed peaks. Guard: Auto-scaling for predictable load variations; manual provisioning for structural growth.
- No Utilization Monitoring: Assume servers fully utilized. Result: under-utilization, wasted cost. Guard: Monitor and right-size; consolidate low-utilization servers.
Further Reading
- Capacity Planning by Neil J. Gunther — quantitative capacity planning
- The Art of Scalability by Martin Abbott and Michael Fisher — growth strategies
- AWS Cost Optimization — rightsizing and utilization analysis