Design Netflix

Problem statement

Design a subscription video-on-demand platform that ingests studio and original content, processes it into protected multi-bitrate streams, publishes a regionally licensed catalog, personalizes a row-based home experience per profile, and plays video on thousands of device types with minimal startup delay and rebuffering. The product must manage memberships and multiple profiles per household, persist watch history and resume points across devices, enforce licensing windows per territory, and support both subscription and ad-supported tiers.

This is not a file-hosting problem. A video platform is dominated by three physical realities: bytes are enormous compared to any other consumer workload, quality of experience is perceived continuously rather than at request time, and the catalog itself is a legal artifact whose availability differs by country and by date. A strong design therefore separates four planes and reasons about each with its own consistency, latency, and cost model.

The four planes

1. Content supply plane: studio delivery of mezzanine masters, validation, transcoding into per-title bitrate ladders, quality control with perceptual metrics, packaging into encrypted segments, dubbing and subtitles, and publish workflow with licensing windows per territory.
2. Delivery plane: a cache-heavy distribution network placed inside ISP networks, request routing, predictive prepositioning of popular assets, adaptive bitrate selection on the client, and DRM license issuance.
3. Experience plane: catalog and search, row-based personalized home, artwork personalization, profiles, watch history, resume points, and device registry.
4. Commerce plane: subscription lifecycle, billing, entitlements, household and concurrency rules, ad-supported tier decisioning, and customer support tooling.

Public operating baseline versus design assumptions

Public evidence establishes the category scale. Netflix reported approximately 302 million paid memberships at the end of 2024 in its shareholder communication, operates in more than 190 countries since its 2016 global expansion, and disclosed roughly 94 million monthly active users on its ad-supported tier in 2025. Netflix's engineering publications describe Open Connect, a cache appliance fleet deployed inside ISP networks and at internet exchange points that serves the large majority of playback traffic, and describe per-title encoding and the VMAF perceptual quality metric that the company open-sourced. Disney+ Hotstar publicly reported 25.3 million concurrent streams for a single cricket match in 2020, demonstrating that live-adjacent concurrency spikes are real for streaming platforms.

For capacity planning this answer explicitly assumes: 45 million concurrent streams at global prime-time peak, 260 million play sessions per day average, a 96 percent cache hit ratio at edge caches, and a 4x evening peak multiplier on session starts. Unless a number is tied to a named public source, it is a stated design assumption, target, or budget.

Why the problem is distinctive

A social feed can retry a request; a video session cannot retry a missed segment without the viewer perceiving a stall. The design therefore treats quality of experience — startup delay, rebuffer ratio, bitrate stability — as the primary service objective, and treats the CDN, the encoder farm, and the client adaptation algorithm as one coupled control system rather than three independent components. Licensing adds a second distinctive constraint: the same asset is simultaneously available and unavailable depending on territory and date, so catalog correctness is a geo-temporal entitlement problem, not a caching problem.