Design Netflix

Netflix builds and operates their own CDN hardware called Open Connect Appliances (OCA). These are custom-built servers (8–250TB of flash storage, 100 Gbps network interfaces) installed inside ISP data centers and exchange points worldwide. When a subscriber in New York watches Stranger Things, the video bytes travel only from their ISP's machine room to their home — never crossing the public internet to AWS. This is why Netflix has near-zero buffering even during peak hours.

Netflix knows which shows are releasing tomorrow (they made them). Every night during off-peak hours, Netflix pushes video chunks for upcoming releases to all relevant Open Connect appliances worldwide. By the time subscribers click "play" on a new season premiere, the files are already on their ISP's hardware. This eliminates origin-server load spikes on launch night. For older content: ML predicts which titles will be watched this week and pre-stages accordingly.

The control plane runs on AWS with 200+ microservices: Authentication, Profile, Catalog, Search, Recommendations, Playback URL generation, Analytics, Billing, A/B Testing, etc. Netflix pioneered Chaos Engineering: deliberately injecting failures into production (Chaos Monkey kills random service instances) to verify resilience. Services communicate via REST and Kafka. All services are deployed across 3 AWS regions (us-east-1, eu-west-1, ap-southeast-1) for failover.

Netflix's recommendation system is its core differentiator — 80% of watched content comes from recommendations (not search). The system uses: (1) Collaborative filtering: "users similar to you also watched X" (matrix factorization on watch history), (2) Content-based filtering: "you liked action thrillers, here are more", (3) Contextual signals: time of day, device type, day of week, (4) A/B testing: different recommendation algorithms run simultaneously for different user cohorts.

Netflix doesn't use the same bitrate for every video. Each title is analyzed frame-by-frame and encoded at the minimum bitrate that maintains quality for that specific content. An animated cartoon needs far less bandwidth than a dark, complex nature documentary at the same visual quality. This "per-title encoding" reduces storage and bandwidth by 20–40%. Netflix uses AV1 (30% more efficient than H.264) for supported devices.

Netflix runs 100s of simultaneous A/B experiments. Every user cohort may see a different: thumbnail image for a show, row ordering on the home screen, recommendation algorithm, UI layout, even video encoding quality level. The experimentation platform randomly assigns users to treatment groups, measures engagement metrics (click-through rate, completion rate, retention), and statistically determines winning variants. Features ship if they win A/B tests with statistical significance.