Case Study: Video Streaming (Netflix/YouTube)
Difficulty: Advanced | Category: Data-Intensive & High Bandwidth | Similar Systems: Twitch, Hulu
Designing a video streaming platform is unique because the challenge isn't just request volume (QPS)—it's raw network bandwidth and massive storage requirements.
1. Requirements Clarification
Functional Requirements
- Upload: Content creators (or internal studios) can upload raw video files.
- Streaming: Users can watch videos on different devices (Mobile, TV, Web) at varying bandwidths.
- Search: Users can search for video titles.
Non-Functional Requirements
- No Buffering: Video must stream smoothly with minimal latency.
- High Availability: Must survive regional data center outages.
- Global Reach: Fast streaming regardless of geographic location.
2. Back-of-the-Envelope Estimation
Assume 200 million users.
- Storage: If 1,000 hours of video are uploaded daily. A 1-hour raw 4K video is ~20 GB.
- 1000 * 20 GB = 20 TB/day of raw video.
- After transcoding into 5 different resolutions/codecs, storage multiplies to ~100 TB/day.
- Bandwidth: If 50 million users watch concurrently during peak hours 3 Mbps (HD quality):
- 50M * 3 Mbps = 150 Terabits per second (Tbps) of egress bandwidth.
WARNING
150 Tbps is massive. A standard central data center cannot serve this. The architecture requires a heavily optimized Content Delivery Network (CDN) pushed to the absolute edge of the network.
3. High-Level Architecture: The Two Planes
Streaming architectures are split into two distinct halves: the Control Plane and the Data Plane.
A. The Control Plane (Backend Services)
Handles everything except streaming the actual video file. Hosted in a central cloud (like AWS).
- User Login & Profiles
- Recommendations & Search
- Metadata Database (Titles, actors, descriptions)
- Video Transcoding Pipeline
B. The Data Plane (Video Streaming)
Handles the heavy lifting of delivering video bytes. Hosted on Edge CDNs (e.g., Netflix's Open Connect Network).
4. Deep Dive: Video Transcoding (DAG Pipeline)
When a raw .mp4 is uploaded, it cannot be streamed directly. It must be processed into smaller chunks and different formats.
- Chunking: The video is split into small chunks (e.g., 5-second segments).
- Transcoding Pipeline: A Directed Acyclic Graph (DAG) of workers processes the chunks.
- Converts to 1080p, 720p, 480p, 360p.
- Applies different codecs (H.264 for older devices, AV1 for modern web).
- Extracts audio tracks and subtitles.
- Packaging: The chunks are wrapped in protocols like HLS (HTTP Live Streaming) or DASH (Dynamic Adaptive Streaming over HTTP).
Why chunking? If a user's Wi-Fi drops in speed, HLS allows the player to seamlessly download the next 5-second chunk at 480p instead of 1080p, preventing buffering (Adaptive Bitrate Streaming).
5. Deep Dive: The Edge CDN Strategy
To serve 150 Tbps globally, you cannot stream from AWS us-east-1 to users in India.
- ISPs (Internet Service Providers): Companies like Netflix install specialized hardware appliances (Open Connect Appliances) directly inside ISP data centers (Comcast, AT&T, Jio).
- Proximity: When you hit "Play", the Control Plane determines your location and ISP and gives your app the IP address of the CDN appliance closest to your house.
- Predictive Caching: CDNs have limited storage. Netflix pushes the most popular shows to edge CDNs overnight during off-peak hours based on regional popularity (e.g., caching popular anime heavily in Tokyo edge nodes).
6. Database Selection
- Video Metadata (Titles, Descriptions): Relational DB (PostgreSQL/MySQL) for ACID compliance and complex queries, cached heavily with Redis.
- Raw & Processed Video Files: Blob/Object Storage (AWS S3) due to cost-effectiveness and durability for massive unstructured files.
- Search: Elasticsearch, syncing data from the Metadata DB to provide fast autocomplete and fuzzy searching.