Case Study: Netflix

Netflix peak time-এ internet-এর ১৫-৩৫% bandwidth use করে। ২০০+ million subscriber, ১৯০+ country, প্রতিদিন billion+ hour viewing। এই scale handle করার engineering — system design-এর masterclass।

Requirements

Functional

• Video streaming (multiple devices, qualities)।
• Browse + search content।
• Personalized recommendation।
• User profile, watch history।
• Subscription, billing।
• Multi-device sync।

Non-Functional

• Low latency, smooth playback।
• High availability (99.99%+)।
• Adaptive quality।
• Global delivery।
• Massive bandwidth।

Capacity Estimation

Architecture Overview

Netflix Open Connect (CDN)

Netflix-এর নিজস্ব CDN — ১০০+ ISP-এর সাথে partnership।

Open Connect Appliance (OCA)

• Server hardware ISP-এর data center-এ install।
• Free for ISP — Netflix থেকে।
• Pre-popular content cache।
• User-এর কাছে network-এ short hop।

কেন কাজ করে

• Bangladesh ISP-এর OCA থাকলে — Netflix content local।
• Latency কম, ISP-এর external bandwidth save।
• Viewing pattern predictable (popular show peak hour-এ)।

Video Encoding Pipeline

একটি 4K movie ingest করতে:

1. Master file (raw)।
2. Encoded multiple bitrate (240p, 480p, 720p, 1080p, 4K)।
3. Each bitrate-এ multiple codec (H.264, H.265, AV1)।
4. Multiple language audio।
5. Subtitle (multi-language)।
6. Each combination separate file → হাজার হাজার version।
7. Encoded files OCA-তে push।

Adaptive Bitrate Streaming (HLS/DASH)

• Video chunks (e.g., 10-second segments)।
• Each chunk multiple quality।
• Player network-অনুযায়ী quality select।
• Network slow → automatically lower quality।

Microservices Architecture

Netflix microservice pioneer — 700+ services।

• Account, billing, subscription।
• Catalog, metadata।
• Search, recommendation।
• Playback, encoding।
• Analytics, A/B testing।

Open-source contributions

• Eureka: Service discovery।
• Hystrix: Circuit breaker (deprecated)।
• Zuul: API Gateway।
• Spinnaker: CD platform।
• Chaos Monkey: Chaos engineering।

Recommendation System

Netflix-এর secret sauce — 80% viewing recommendation থেকে।

Approaches

• Collaborative filtering: Similar user-এর preference।
• Content-based: Movie attribute (genre, actor) match।
• Deep learning: Neural network।
• Contextual: Time of day, device, mood।

Pipeline

1. User behavior log (Kafka)।
2. Batch processing (Spark) — model retrain।
3. Real-time scoring।
4. Cache top recommendations।
5. A/B test continuously।

Playback Flow

1. User movie click।
2. Playback service consult — license check, geo-restriction।
3. Manifest file (DASH/HLS) generate — chunk URLs সহ।
4. Player nearest OCA-এর address পায়।
5. Player chunks download — adaptive bitrate।
6. Buffer maintain।
7. Watch position periodically save।

Cloud-Native (AWS)

Netflix all-in on AWS:

• EC2 — compute।
• S3 — master storage।
• DynamoDB — state।
• Cassandra — viewing history (custom multi-region)।
• EMR — data processing।

Multi-region active-active।

Chaos Engineering

Netflix Chaos Monkey + suite:

• Chaos Monkey: Random instance kill।
• Latency Monkey: Inject latency।
• Chaos Kong: Whole region fail।
• Goal: Production resilience verify।

A/B Testing

Everything A/B test:

• Thumbnail image।
• Title placement।
• Recommendation algorithm।
• UI element।

Data-driven decision।

Trade-offs

• Microservice complexity vs scalability।
• Encoding storage vs adaptive quality।
• Recommendation accuracy vs compute cost।
• OCA hardware cost vs ISP partnership benefit।

Engineering Lessons

1. Own your CDN if you can।
2. Embrace failure (chaos engineering)।
3. Microservice + DevOps maturity।
4. Data-driven (A/B everything)।
5. Cloud-native, multi-region।