Design a Video Streaming Service (YouTube/Netflix)

Build a global video platform with encoding pipelines, CDN distribution, adaptive streaming, and recommendation systems for billions of hours of content.

System Requirements

Functional Requirements

Upload videos (multiple formats, up to 4K resolution)
Transcode to multiple bitrates and formats
Global video streaming with adaptive bitrate
Video search and discovery
User profiles and video management
Comments, likes, and social features
Live streaming capabilities
Content recommendations based on viewing history

Non-Functional Requirements

2B hours watched per day (YouTube scale)
500 hours of video uploaded per minute
Support 4K, 1080p, 720p, 480p, 360p streaming
Video start time < 2 seconds globally
Rebuffering ratio < 0.5%
99.9% availability for streaming
Support 50M+ concurrent viewers
Global CDN with <200ms latency

Capacity Estimation

Video Platform Scale

Read vs Write

1x ratioUpload

1000x ratioWatch

Content Type

5% splitLive

95% splitVOD

Quality Distribution

30% viewsSD

70% viewsHD+

Traffic Analysis

Daily Watch Hours

YouTube-scale consumption

2B hours

Upload Rate

26,000 hours uploaded daily

500 hrs/min

Peak Concurrent

Global prime time overlap

50M users

Bandwidth Peak

Global aggregate throughput

40 Tbps

Storage Requirements

Raw Video Storage

100+ petabytes total content

Transcoded Versions

5x storage multiplication factor

Daily Growth

1 petabyte new content/day

Video Encoding Pipeline

Upload Path:

Client → Upload Service → Object Storage (Raw Video)

↓

Transcoding Queue → Worker Nodes (FFmpeg/GPU)

↓

Multiple Bitrates/Formats → CDN Origin Servers

↓

Metadata Extraction → Search Index + Recommendations

Streaming Path:

Client Request → CDN Edge → Adaptive Bitrate Selection

↓

Video Segments (HLS/DASH) → Progressive Download

↓

Player Buffer Management → Quality Adaptation

Video Format Matrix

Resolution	Bitrate	Use Case	Storage Factor	Complexity
4K (2160p)	35-45 Mbps	Premium content, high-end devices	1x (reference)	Very High
1080p HD	8-12 Mbps	Standard HD viewing	0.3x	High
720p	4-6 Mbps	Mobile devices, medium bandwidth	0.15x	Medium
480p	2-3 Mbps	Low bandwidth connections	0.08x	Low
360p	1-1.5 Mbps	Very low bandwidth, mobile	0.05x	Low

System Architecture

Upload Service

Multipart upload, S3/GCS, chunked transfer

Purpose:

Handle video file uploads with resumable uploads

Scale:

500 hours uploaded/minute

Transcoding Pipeline

FFmpeg, Kubernetes jobs, GPU acceleration

Purpose:

Convert videos to multiple formats and bitrates

Scale:

1M+ hours processed/day

CDN Network

CloudFront/Cloudflare, adaptive bitrate

Purpose:

Global video delivery with edge caching

Scale:

200+ edge locations worldwide

Metadata Service

Elasticsearch, ML recommendation engine

Purpose:

Video metadata, search indexing, recommendations

Scale:

100B+ video views/day

Upload Tier

• Resumable multipart uploads

• Virus scanning & validation

• Duplicate detection

• Content moderation queue

Processing Tier

• Horizontal transcoding cluster

• Priority queue management

• GPU-accelerated encoding

• Quality assurance checks

Delivery Tier

• Global CDN distribution

• Edge caching optimization

• Adaptive bitrate streaming

• Analytics & monitoring

Streaming Optimizations

Adaptive Bitrate Streaming (ABR)

Dynamically adjust video quality based on network conditions

Benefit:

60% reduction in rebuffering events

Implementation:

HLS/DASH protocols, client-side bandwidth estimation

Predictive Caching

Pre-cache popular content at edge locations

Benefit:

80% cache hit rate for trending videos

Implementation:

ML-based popularity prediction, proactive cache warming

Smart Thumbnails

Generate multiple thumbnail options and A/B test

Benefit:

20% increase in click-through rates

Implementation:

Computer vision for scene detection, user engagement analysis

Progressive Download

Start playback while continuing to download

Benefit:

Sub-2 second video start time

Implementation:

Segment-based streaming, buffer optimization

Performance Metrics

Video Quality Metrics

Video Start Time

Time to first frame globally

< 2 seconds

Rebuffering Ratio

Industry-leading performance

< 0.5%

Quality Adaptation

Network change response time

< 500ms

CDN Cache Hit Rate

Encoding Performance

Encoding Speed

For 1080p content with GPUs

10x real-time

Processing Queue

Average wait time for uploads

< 30 minutes

Storage Efficiency

Advanced codecs vs H.264

60% savings

Parallel Jobs

Peak transcoding capacity

10K concurrent

Technical Challenges & Solutions

Processing Speed vs Quality

Problem: Balance between encoding speed and video quality at scale

Solution: Multi-pass encoding with quality-based rate control

Implementation: Hardware acceleration (GPU/ASIC), parallel processing

Format Compatibility

Problem: Support diverse devices and browsers with different codecs

Solution: Multiple codec support: H.264, H.265, VP9, AV1

Implementation: Progressive codec adoption, fallback mechanisms

Storage Optimization

Problem: Massive storage requirements for multiple quality versions

Solution: Intelligent tiering, compression optimization, cold storage

Implementation: S3 Intelligent Tiering, automated lifecycle policies

Real-time Processing

Problem: Live streaming requires real-time low-latency transcoding

Solution: Dedicated live transcoding pipeline with chunked processing

Implementation: WebRTC ingestion, low-latency HLS/DASH streaming

Database Design

Video Metadata

videos table:
- video_id (UUID, PK)
- user_id (UUID, indexed)
- title (varchar, full-text indexed)
- description (text)
- tags (JSON array)
- duration (seconds)
- upload_date (timestamp)
- view_count (bigint, default 0)
- like_count (int, default 0)
- status (processing/ready/error)
- thumbnail_url (varchar)
- privacy_setting (public/private/unlisted)

Video Files

video_files table:
- file_id (UUID, PK)
- video_id (UUID, FK)
- resolution (varchar) 
- bitrate (int)
- format (varchar)
- codec (varchar)
- file_size (bigint)
- storage_path (varchar)
- created_at (timestamp)

Index: (video_id, resolution) for ABR

Analytics Schema

video_analytics table (time-series):
- event_id (UUID, PK)
- video_id (UUID, indexed)
- user_id (UUID, nullable)
- event_type (view/like/share/comment)
- timestamp (timestamp)
- session_id (UUID)
- device_type (mobile/desktop/tv)
- geo_location (varchar)
- watch_duration (int, for view events)

Partitioned by date for efficient queries

Practice Questions

Design a fault-tolerant transcoding pipeline. How do you handle failed jobs and ensure all videos are processed?

How would you implement adaptive bitrate streaming? Design the client-side logic for quality switching.

Design a recommendation system for video content. What signals would you use and how would you scale it?

How would you handle live streaming? What are the differences in architecture compared to video-on-demand?

Design a content moderation system. How do you automatically detect and flag inappropriate content at scale?