Estimate infrastructure or operational capacity needed for Recommendations at global scale

Problem Statement Description

Product context: Netflix is a streaming entertainment company; its products include subscription video, original films and series, recommendations, profiles, games, and ad-supported plans.

Netflix’s Recommendations experience helps members decide what to watch across a global catalog that includes films, series, live events, games, ads-supported experiences, and locally relevant content. In this guesstimate, you are asked to estimate the infrastructure or operational capacity required to support Recommendations at global scale, with particular attention to casual viewers who may browse briefly, switch devices, and rely heavily on personalized rows, thumbnails, search-adjacent discovery, and “continue watching” cues.

Frame the problem as a capacity-estimation exercise rather than a product redesign. You should define what “Recommendations capacity” includes, such as request volume, model inference, ranking, feature retrieval, cache usage, data refreshes, experimentation load, and operational support. You may choose a reasonable scope, but make it explicit: for example, whether you are estimating real-time serving capacity, offline training/data pipelines, human operations, or an end-to-end blended view.

Your answer should show how you break a vague global-scale infrastructure question into measurable units, make assumptions, pressure-test them, and identify the variables that most affect the estimate. The interviewer is looking for structured reasoning, not exact Netflix internal numbers.

The experience should consider:

- The population and usage scope: global subscribers, active profiles, casual viewers, households, devices, and regions.

- The recommendation surfaces included: home page rows, title detail pages, search support, notifications, ads-supported contexts, games, live content, or localized collections.

- The unit of capacity being estimated: recommendation requests per second, daily inference calls, ranking computations, storage, bandwidth, compute clusters, cache footprint, data pipeline jobs, or operational staffing.

- Frequency and seasonality: daily sessions, browse events per session, peak-hour concurrency, weekend spikes, new-release spikes, live-event surges, and regional time-zone effects.

- Assumptions around personalization depth: number of candidate titles, ranking passes, model complexity, feature lookups, thumbnail personalization, A/B tests, and fallback logic.

- Sensitivity drivers: subscriber growth, catalog size, localization needs, latency targets, cache hit rate, model refresh cadence, device mix, and experimentation volume.

- Sanity checks: compare peak versus average demand, online serving versus offline processing, global versus regional load, and whether estimates align with a consumer internet service operating at Netflix scale.

The goal is to produce a clear, defensible estimate of the infrastructure or operational capacity needed to keep Netflix Recommendations fast, personalized, localized, and reliable worldwide, while clearly stating assumptions, boundaries, and the factors that would change the estimate most.