Netflix

Global video streaming platform serving 260M+ subscribers with personalized content delivery, adaptive bitrate streaming, and microservices at massive scale.

260M+

subscribers

~500M hours/day

daily Streaming Hours

~15% of global internet downstream traffic

bandwidth

1,000+ microservices

microservices

~1,200 variants per title

encoding Variants

17,000+ servers in 6,000+ locations

open Connect Servers

Architecture Diagram

Data Flow

Client App → API GatewayBrowse

Client requests homepage. Gateway routes to Content API with auth context.

API Gateway → Content APIFetch Content

Content API assembles personalized rows by calling Recommendation Engine and Catalog DB.

Content API → EVCacheCache Check

Hot data served from EVCache. Cache miss falls through to Catalog DB.

Content API → Recommendation EnginePersonalize

ML models rank content based on viewing history, time of day, device, and A/B test group.

Client App → API GatewayPlay Request

User clicks play. Client sends play request with device capabilities.

API Gateway → Playback ServiceGet Manifest

Playback service generates DASH/HLS manifest with DRM license and optimal CDN URLs.

Client App → Open Connect CDNStream Video

Client fetches video segments from nearest Open Connect appliance. Adaptive bitrate adjusts quality.

Client App → KafkaViewing Events

Client streams viewing telemetry (play, pause, seek, quality changes) to Kafka for analytics.

Encoding Pipeline → Object StoreStore Encoded

New content is transcoded into multiple bitrates/resolutions and stored in S3.

Object Store → Open Connect CDNPre-position

Popular content is pre-positioned to ISP appliances during off-peak hours.

Key Architectural Decisions

Custom CDN (Open Connect) placed inside ISPs reduces transit costs and latency — 95% of traffic never crosses the internet backbone
Per-title encoding optimizes bitrate ladders — an animated show needs fewer bits than live action at the same perceptual quality
Microservices with Zuul gateway enables independent deployment of 1,000+ services by hundreds of teams
EVCache (memcached-based) chosen over Redis for simplicity at Netflix's specific scale and access patterns
Chaos engineering (Chaos Monkey) as a first-class practice ensures resilience is tested continuously

Tradeoffs

Strengths

Open Connect CDN is a massive competitive moat — better quality at lower cost than any third-party CDN
Per-title encoding delivers optimal quality per bit, reducing bandwidth costs
Microservices architecture enables Netflix's rapid feature velocity across many teams
Chaos engineering culture means the system is tested for failure continuously

Weaknesses

Microservices complexity requires massive investment in platform tooling (service mesh, observability)
Open Connect requires physical hardware relationships with ISPs globally
Recommendation system complexity means debugging "why was this recommended" is extremely difficult
Cold start problem for new users with no viewing history

Interview Drilldown Questions

How does adaptive bitrate streaming work, and what are the tradeoffs between DASH and HLS?
How would you design the content recommendation system to handle 260M users?
What's the strategy for pre-positioning content on CDN nodes?
How does Netflix handle regional content licensing in its catalog?
How would you design the A/B testing infrastructure at Netflix's scale?