Case Studies/Streaming

Netflix

Global video streaming with personalized content delivery

streamingvideoCDNmicroservicesrecommendation

Case study boundary

Use this as a learning model.

This page explains architectural patterns and constraints inspired by Netflix. It should not be read as the exact private production architecture or as a one-size-fits-all recommendation.

Evidence

Public info + inference

Learn first

Concepts behind this architecture

Overview

Netflix is the world’s leading streaming service with over 260 million paid subscribers in 190+ countries. Its architecture is a masterclass in cloud-native design, having pioneered many microservices and chaos engineering patterns. Netflix serves over 1 billion hours of video per week through its custom CDN (Open Connect), while its recommendation engine drives 80% of content watched.

Scale

260M+

Paid subscribers

1B+ hours

Video streamed weekly

~500K

API requests/second

Petabytes

Content encodings

Requirements

Functional

  • Video streaming with adaptive bitrate
  • Personalized recommendations and search
  • User profiles and viewing history
  • Offline downloads
  • Multiple concurrent streams per account
  • Content browsing across genres

Non-Functional

  • Global low-latency video delivery
  • 99.99% availability
  • Adaptive quality based on network conditions
  • Support for 100+ device types
  • Horizontal scalability for peak traffic
  • Fault tolerance — no single point of failure

High-Level Architecture

Loading diagram...
Netflix — High-level system architecture

Key Components

API Gateway (Zuul)

Edge service handling request routing, authentication, rate limiting, and load shedding. Processes billions of requests per day with dynamic routing rules.

Open Connect CDN

Netflix’s custom CDN with thousands of appliances embedded in ISP networks worldwide. Serves 95%+ of video traffic, reducing internet backbone load.

Playback Service

Determines the optimal video stream for each client based on device capabilities, network conditions, and DRM requirements.

Recommendation Service

ML-powered personalization using collaborative filtering, content-based filtering, and deep learning. Powers 80% of what users watch.

Encoding Pipeline

Transcodes each title into 100+ encodings (resolution, bitrate, codec) using per-title and per-shot optimization for quality-efficient streaming.

EVCache

Distributed caching layer (built on Memcached), handling 30M+ requests/second with sub-millisecond latency for session, profile, and recommendation data.

Data Model

Loading diagram...
Netflix — Entity relationship diagram

Open Connect CDN Architecture

Netflix’s Open Connect is one of the largest CDNs in the world, purpose-built for video delivery.

How it works:

  • Netflix places custom server appliances (OCAs) inside ISP networks worldwide
  • During off-peak hours, new content is proactively pushed to OCAs based on predicted demand
  • When a user presses play, they’re directed to the closest OCA that has the content

OCA Specifications:

  • Custom-built servers with 100+ TB SSD storage
  • Optimized FreeBSD-based OS with custom TCP stack
  • Single OCA can serve 100+ Gbps of video
  • ~18,000 OCAs in 6,000+ locations worldwide

Content Placement Algorithm:

  • Predicts what content will be popular in each region
  • Fills OCAs during off-peak hours (typically 2am-8am local time)
  • Popular content replicated across more OCAs for redundancy

95% of Netflix traffic never crosses the internet backbone, reducing costs and improving quality.

Loading diagram...
Open Connect CDN Architecture

Adaptive Bitrate Streaming

Netflix uses adaptive bitrate (ABR) streaming to deliver the best quality for each user’s network conditions.

Per-Title Encoding: Instead of fixed encoding ladders, Netflix analyzes each title’s visual complexity and creates a custom encoding ladder:

  • Simple animation might look perfect at 1.5 Mbps in 1080p
  • An action movie might need 8 Mbps for the same resolution
  • Saves 20% bandwidth on average while maintaining quality

Per-Shot Encoding:

  • Static dialogue scenes get lower bitrates
  • Fast-action sequences get higher bitrates
  • Smooth transitions between quality levels

Client-Side ABR Algorithm:

  1. Maintains a ~30-second playback buffer
  2. Measures download speed of recent segments
  3. Selects the highest sustainable quality
  4. Upgrades/downgrades smoothly between segments

Netflix encodes each title into ~120 different streams (resolution, bitrate, codec, HDR combinations).

Chaos Engineering & Resilience

Netflix pioneered chaos engineering — intentionally introducing failures to build system resilience.

Key Tools:

  • Chaos Monkey: Randomly terminates production instances
  • Chaos Kong: Simulates entire AWS region failure
  • Latency Monkey: Introduces artificial delays

Resilience Patterns:

  • Circuit Breakers: Calls to failing services are short-circuited with fallbacks
  • Bulkheads: Thread pool isolation prevents cascade failures
  • Fallbacks: Every service has degraded-mode responses
  • Retries with backoff: Exponential backoff with jitter

The philosophy: “The best way to avoid failure is to fail constantly.” By regularly testing failure scenarios in production, Netflix ensures graceful handling of real failures.

Architectural Tradeoffs

Custom CDN (Open Connect) vs. third-party CDN

Pros

  • +Full control over hardware/software stack
  • +Massive cost savings at scale
  • +95%+ traffic served from ISP edge

Cons

  • −Huge upfront infrastructure investment
  • −Complex ISP relationships
  • −Hardware logistics across 6000+ locations

Microservices architecture

Pros

  • +Independent deployability
  • +Technology heterogeneity
  • +Fault isolation between services

Cons

  • −Distributed system complexity
  • −Network latency between services
  • −Harder debugging across boundaries

Per-title encoding over fixed ladders

Pros

  • +20% bandwidth savings
  • +Better quality for simple content
  • +Optimal experience per title

Cons

  • −Much higher encoding compute cost
  • −Longer content onboarding time
  • −Complex encoding pipeline

Learn From Netflix

Copy a learning context prompt for an AI assistant. It keeps assumptions visible and asks for tradeoffs, cost drivers, failure modes, and what to measure before adding complexity.

Explore More Case Studies