Video Streaming System Design (Netflix/YouTube)

Overview

Design a scalable video streaming platform capable of serving millions of concurrent users with high-quality video content, adaptive bitrate streaming, and personalized recommendations.

Requirements

Functional Requirements

• Upload and store videos in multiple formats and qualities
• Stream videos with adaptive bitrate (ABR) support
• Search and browse video catalog
• User profiles and watch history
• Video recommendations
• Playback controls (play, pause, seek, quality selection)
• Subtitle support
• Like, comment, and share functionality

Non-Functional Requirements

• Low latency video streaming (<2s startup time)
• High availability (99.99% uptime)
• Support 100M+ concurrent users
• Global content delivery
• Minimal buffering
• Cost-effective storage and bandwidth

Capacity Estimation

Storage

• Average video size: 1GB (compressed)
• 1 million videos uploaded daily
• Daily storage: 1TB/day
• 5 years: 1TB × 365 × 5 = 1.8PB
• Multiple qualities: 1.8PB × 5 = 9PB

Bandwidth

• 100M concurrent viewers
• Average bitrate: 5 Mbps
• Total bandwidth: 500 Tbps peak
• With CDN caching: ~50 Tbps from origin

Traffic

• Daily video views: 5 billion
• Average video length: 10 minutes
• Total watch time: 50 billion minutes/day

High-Level Design

Components

1. Video Upload Service

   • Handle video file uploads
   • Validate and quarantine content
   • Queue for processing

2. Video Processing Pipeline

   • Transcoding to multiple formats (MP4, WebM, HLS)
   • Generate multiple quality levels (360p, 480p, 720p, 1080p, 4K)
   • Thumbnail generation
   • Subtitle extraction/encoding
   • Video metadata extraction

3. Content Delivery Network (CDN)

   • Global edge servers
   • Cache popular content
   • Geo-distributed storage
   • Origin shield

4. Video Storage

   • Object storage (S3, GCS)
   • Multiple redundancy zones
   • Hot/warm/cold storage tiers

5. Streaming Service

   • Adaptive bitrate streaming (HLS/DASH)
   • DRM and encryption
   • Analytics tracking

6. Recommendation Engine

   • Machine learning models
   • Collaborative filtering
   • User behavior analysis

7. API Gateway

   • Authentication/authorization
   • Rate limiting
   • Request routing

8. Metadata Service

   • Video information
   • User profiles
   • Watch history
   • Recommendations

9. Search Service

   • Elasticsearch/Solr
   • Auto-complete
   • Fuzzy matching

Database Schema

Videos Table

sql
{
  "videoId": "v123",
  "title": "Sample Video",
  "description": "Video description",
  "uploaderId": "user456",
  "uploadDate": "2024-01-01T00:00:00Z",
  "duration": 600,
  "views": 1000000,
  "likes": 50000,
  "categories": ["education", "technology"],
  "thumbnailUrl": "https://cdn.example.com/thumb.jpg",
  "status": "published"
}

Video Files Table

sql
{
  "videoId": "v123",
  "quality": "1080p",
  "format": "mp4",
  "url": "https://cdn.example.com/v123/1080p.mp4",
  "size": 1073741824,
  "bitrate": 5000000,
  "codec": "H.264"
}

Watch History Table

sql
{
  "userId": "user456",
  "videoId": "v123",
  "watchedAt": "2024-01-01T12:00:00Z",
  "duration": 300,
  "completed": false,
  "lastPosition": 300
}

Video Processing Workflow

1. Upload

   • User uploads video
   • Store in temporary storage
   • Generate upload ID
   • Return success

2. Validation

   • Scan for malware
   • Check copyright (Content ID)
   • Validate format and codec
   • Check file integrity

3. Transcoding

   • Parallel processing for multiple qualities
   • Use GPU-accelerated encoders
   • Generate adaptive bitrate manifests
   • Extract audio tracks

4. Storage

   • Upload to CDN origin
   • Replicate across regions
   • Update metadata service

5. Publish

   • Update video status
   • Trigger indexing
   • Send notifications

Streaming Protocol

Adaptive Bitrate Streaming (HLS/DASH)

text
GET /video/v123/manifest.m3u8

#EXTM3U
#EXT-X-STREAM-INF:BANDWIDTH=5000000,RESOLUTION=1920x1080
1080p.m3u8
#EXT-X-STREAM-INF:BANDWIDTH=2500000,RESOLUTION=1280x720
720p.m3u8
#EXT-X-STREAM-INF:BANDWIDTH=1000000,RESOLUTION=854x480
480p.m3u8

CDN Architecture

Edge Locations

• 200+ PoPs worldwide
• Cache popular content at edge
• Origin shield to reduce load
• Regional clusters

Caching Strategy

• Hot content: Cache at all edges (trending videos)
• Warm content: Cache at regional level (recent uploads)
• Cold content: Serve from origin (old/unpopular videos)
• TTL-based eviction
• Predictive caching based on trends

Scaling Considerations

Horizontal Scaling

• Microservices architecture
• Containerization (Docker/Kubernetes)
• Auto-scaling based on load
• Load balancers at each tier

Database Sharding

• Shard by videoId (hash-based)
• Separate read replicas for analytics
• Write-heavy: Cassandra/DynamoDB
• Read-heavy: Redis/Memcached

Video Processing

• Distributed transcoding cluster
• Queue-based processing (Kafka/RabbitMQ)
• Spot instances for cost optimization
• Priority queues for paid users

Cost Optimization

Storage Tiers

• Hot tier: Frequently accessed (SSD)
• Warm tier: Occasionally accessed (HDD)
• Cold tier: Rarely accessed (Glacier)
• Automatic lifecycle policies

Bandwidth Optimization

• Aggressive CDN caching
• P2P streaming for live events
• Video compression (H.265, AV1)
• Smart quality selection based on network

Computing

• Serverless for API gateway
• Spot instances for transcoding
• Reserved instances for core services

Security & DRM

1. Content Protection

   • Widevine, FairPlay, PlayReady
   • Token-based authentication
   • Encrypted streaming (HTTPS)

2. Copyright Protection

   • Content ID system
   • Fingerprinting
   • Automated takedowns

3. User Security

   • OAuth 2.0 authentication
   • JWT tokens
   • CSRF protection

Monitoring & Analytics

Real-time Metrics

• Video buffering ratio
• Startup time (Time to First Byte)
• Concurrent viewers
• CDN hit ratio
• Error rates by region

Business Metrics

• Watch time
• Engagement rate
• Conversion rate
• Revenue per user

Player Analytics

• Quality switches
• Rebuffering events
• Drop-off points
• Device/browser distribution

API Design

Upload Video

text
POST /api/v1/videos/upload
Headers: Authorization: Bearer <token>
Request: multipart/form-data
Response: {
  "videoId": "v123",
  "uploadUrl": "https://upload.example.com/v123",
  "status": "processing"
}

Get Video Stream

text
GET /api/v1/videos/{videoId}/stream
Response: {
  "manifestUrl": "https://cdn.example.com/v123/manifest.m3u8",
  "thumbnails": [...],
  "subtitles": [...]
}

Search Videos

text
GET /api/v1/search?q=tutorial&limit=20
Response: {
  "results": [...],
  "total": 1000,
  "page": 1
}

Trade-offs

💡 Interview Tips & Out-of-the-Box Thinking

Common Pitfalls to Avoid

• Don't ignore the upload pipeline: Many candidates focus only on playback, but upload optimization (chunked uploads, resumable uploads) is equally critical
• Underestimating storage costs: Storing multiple quality versions of billions of videos requires petabytes of storage - discuss cost optimization strategies
• Forgetting about copyright: Content ID and DRM are crucial for production systems but often overlooked

Advanced Considerations

• Global latency: Discuss edge caching strategies and how to serve users <100ms away from content
• Adaptive bitrate streaming: Explain HLS/DASH and how they dynamically adjust quality based on network conditions
• Live streaming complexity: Live has different requirements (low latency, real-time transcoding, chat sync)
• Monetization: Ad insertion (VAST/VPAID), premium subscriptions, pay-per-view

Creative Solutions

• Predictive pre-caching: Use ML to predict what users will watch next and pre-position content in CDN edge nodes
• Peer-to-peer assistance: For viral videos, use WebRTC to offload some CDN traffic via peer-assisted delivery
• Smart thumbnail generation: Use AI to detect engaging moments in videos for thumbnails (faces, action scenes)
• Regional content popularity: Don't replicate rarely-watched content globally - use demand-driven replication

Trade-off Discussions

• Consistency vs Availability: Video metadata (views, likes) can be eventually consistent, but payment transactions must be strongly consistent
• Push vs Pull CDN: Push gives you control but requires managing cache invalidation; Pull is simpler but slower for first viewer
• Real-time vs Batch: Live transcoding costs 10x more but batch processing can take hours - discuss hybrid approaches

Edge Cases to Mention

• Simultaneous viral traffic: What if 100M users request same video at once? (Answer: CDN distribution, origin shield, rate limiting)
• Copyright strikes during upload: How to prevent pirated content? (Answer: Content fingerprinting, automated scanning)
• Network switches during playback: User moves from WiFi to cellular - how to maintain smooth playback? (Answer: Adaptive bitrate + buffer strategy)
• Storage failures: What if a datacenter storing original videos goes down? (Answer: Multi-region replication, erasure coding)

Advanced Features

Live Streaming

• Low latency protocols (WebRTC, RTMP)
• Real-time transcoding
• Chat integration
• DVR functionality

Offline Viewing

• Download management
• Encryption for downloads
• Expiration policies
• Storage quota management

Smart Recommendations

• Collaborative filtering
• Content-based filtering
• Deep learning models
• A/B testing

Follow-up Questions

1. How would you handle live streaming with <3s latency?
2. How to reduce storage costs for rarely watched videos?
3. How to implement a fair Content ID system?
4. How to handle sudden viral video traffic spikes?
5. How would you implement seek-ahead buffering?