ArchitectAI resilience reviews are here. See what's new
Back to all articles
High Level Design Oct 18, 2026 13 min read

Designing Netflix: Content Delivery, Dynamic Encoding & Open Connect CDN

How Netflix delivers 15% of global internet traffic. An architectural exploration of custom Open Connect appliances, dynamic video encoding, and recommendation graphs.

HLDStreamingOpen ConnectCDNRecommendation Engine
HFHLDForge Architecture TeamStaff Media Systems Architect

1. Scale & Traffic Characteristics

Netflix accounts for a massive percentage of peak global downstream traffic. Rather than relying solely on commercial CDNs like Akamai or CloudFront, Netflix built its own custom CDN called Open Connect.

Open Connect Appliances (OCAs)—custom FreeBSD storage servers with high-speed 100GbE NICs—are embedded directly inside thousands of Internet Service Provider (ISP) networks around the globe. This brings 95%+ of video bytes within single-hop distance of subscribers.

2. Control Plane (AWS) vs Data Plane (Open Connect)

Netflix bifurcates its architecture into two distinct domains: The Control Plane (running in AWS across multiple regions) and the Data Plane (Open Connect OCAs).

The AWS tier manages user signup, billing, catalog browsing, and recommendation generation. The OCA network handles all heavy video byte delivery. Even if AWS suffers a control plane disruption, users with active playback sessions continue streaming uninterrupted from local OCAs.

3. Per-Title & Per-Shot Dynamic Video Encoding

Traditional video pipelines encoded all titles at static bitrates (e.g. 1080p at 5Mbps). Netflix pioneered shot-based encoding: an animated cartoon scene requires far fewer bits than high-action sports with rapid camera pans.

By analyzing complexity at the individual shot level, video quality is maximized while reducing overall bandwidth consumption by 20-30%.

4. Personalized Recommendation System

The recommendation engine combines collaborative filtering, deep reinforcement learning, and contextual bandits. Candidate generation narrows down millions of titles to a few hundred candidates, which a neural ranking model scores according to viewing history, device type, and time of day.

Interactive Modeling ArchitectAI

Model this architecture on the HLDForge Canvas

Turn static box diagrams into living, collaborative systems. Drag and drop microservices, Kafka clusters, and cache tiers. Run automated SPOF detection and STRIDE threat modeling in real time.

HF

Written by HLDForge Architecture Team

Staff Media Systems Architect

The HLDForge Architecture Team publishes practical engineering guides, real-world case studies, and high-concurrency blueprints to help engineers build scalable and fault-tolerant software.

Free forever plan

Ready to master distributed system design?

HLDForge gives you the tools, templates, and AI reviews to design systems like a Staff Engineer. Get started free.

SOC 2 Type II · ISO 27001 aligned · SSO on every paid plan