Forecast Storage Demand Across 100 Data Centers

Design a forecasting approach for daily data-center storage demand. Explain how you would begin with one data center, extend the approach to 100 data centers, and produce forecasts for a newly established data center with little or no history. ### Part 1 — Forecast one data center Define the target and forecasting horizon, then describe the data, baseline, model, and evaluation method. #### What This Part Should Cover - The distinction between new storage demand, net growth, and occupied capacity. - Seasonality, known future information, and chronological evaluation. - Uncertainty and the consequences of under- or overprediction. ### Part 2 — Scale to 100 data centers How would you share information and operate forecasting across the fleet while preserving differences between sites? #### What This Part Should Cover - Local versus shared models, site features, and model-management costs. - Evaluation at individual-site and fleet levels. ### Part 3 — Handle a new data center How would you forecast demand before a new site has enough observations to learn its own pattern? #### What This Part Should Cover - Comparable sites, planned workload assignments, and uncertainty. - Updating the forecast as local observations arrive. ### What a Strong Answer Covers - A forecasting target connected to a concrete capacity-planning decision. - A scalable modeling and data pipeline with honest treatment of missing history. - Protection against leakage and misleading aggregate accuracy. ### Follow-up Questions - How could a capacity limit make observed storage growth understate actual demand? - How would you check whether fleet-level forecasts and site-level forecasts are consistent?