Design YouTube (Video Platform)
DifficultyHard | HelloInterview: problem breakdown
Problem Statement
Design a video streaming platform where users can upload videos, and viewers can watch them with adaptive bitrate streaming. Handle the full pipeline: upload → processing → storage → streaming.
📖Real-world: A fun fact that captures the scale: "Gangnam Style" (2012) broke YouTube's view counter — it was a signed 32-bit integer, capped at ~2.1 billion, and the video blew past it, forcing Google to move to 64-bit counts. That's the whole "approximate counting at scale" theme in one anecdote. The core of this design is the transcoding pipeline — one uploaded video is fanned out into many resolutions/codecs (the classic async queue + worker-per-format pattern), chunked for adaptive bitrate streaming (HLS/DASH), and pushed to a global CDN where the real bytes are served. Security/trust dimensions to raise: Content ID (automated copyright matching against a reference fingerprint database) and upload moderation (CSAM/abuse scanning) are first-class subsystems — at YouTube's "500 hours uploaded per minute," moderation has to be automated pipeline infrastructure, not human review alone.
Requirements
Functional
- Upload video (any format, up to 100GB)
- Transcode video to multiple resolutions (360p, 720p, 1080p, 4K)
- Stream video to users (adaptive bitrate)
- Video search, recommendations
- Comments, likes, view counts
Non-Functional
- 2B users, 500M DAU
- 500 hours of video uploaded per minute
- 1B hours watched per day → ~700M simultaneous streams
- Upload success even with poor network (resumable)
- Low buffering: < 200ms startup latency, < 1% rebuffer rate
- Global availability
Core Design
Upload Pipeline
Client → Upload Service → Object Store (raw video)
↓
Message Queue (Kafka)
↓
Transcoding Workers (FFmpeg)
↓
┌───────────────────────────┐
│ 360p.mp4 │
│ 720p.mp4 │→ CDN Origin Store (S3)
│ 1080p.mp4 │
│ thumbnail.jpg │
└───────────────────────────┘
↓
Update DB: video status = READY
↓
Notify uploader (email/push)Resumable Uploads (critical for large files):
- Client uses chunked upload (e.g., 5MB chunks)
- Server issues an upload session ID
- Client uploads chunks; server tracks which chunks received
- On failure, client resumes from last successful chunk
- Implementation: AWS Multipart Upload, Google Resumable Upload, or TUS protocol
Why async transcodingvideo processing is CPU-intensive, takes minutes. Don't block the HTTP upload response — queue the job and process asynchronously.
Adaptive Bitrate Streaming (ABR)
Video is divided into small segments (2-10 seconds each) at multiple bitrates. A manifest file (.m3u8 for HLS, .mpd for DASH) lists all segments and bitrates.
video.m3u8:
#EXT-X-STREAM-INF:BANDWIDTH=800000,RESOLUTION=640x360
360p/segment_%d.ts
#EXT-X-STREAM-INF:BANDWIDTH=2800000,RESOLUTION=1280x720
720p/segment_%d.ts
#EXT-X-STREAM-INF:BANDWIDTH=5000000,RESOLUTION=1920x1080
1080p/segment_%d.tsClient player monitors download speed → switches to higher/lower bitrate segment URLs dynamically → smooth playback under variable network conditions.
Architecture
Upload Service
(chunked, resumable)
|
Raw Video Store (S3)
|
Message Queue (Kafka)
|
┌──────────┴──────────────┐
Transcoding Workers Thumbnail Workers
(auto-scaling, GPU-enabled)
|
Processed Video Store (S3)
|
CDN Network (CloudFront, Akamai)
|
User's Video Player
Metadata DB (PostgreSQL/Spanner)
- Video metadata, status
- User profiles, subscriptions
Search (Elasticsearch)
- Video title, description, tags
Recommendations (ML pipeline - Kafka → Flink → serving)Transcoding at Scale
- Videos are independent; parallelizable: spin up many workers
- Each worker picks one job from queue (Kafka consumer group)
- Worker downloads raw video from S3, runs
ffmpeg, uploads segments to S3 - Worker marks job complete in DB; updates video status
For very long videos: segment-level parallelism — split video into 10-minute chunks, transcode in parallel, merge.
CDN Strategy
Videos are served via CDN. CDN caches video segments at edge nodes close to users.
- Popular videos: hot in CDN cache → very low origin load
- Long-tail videos: cache miss → CDN fetches from S3 → caches for next request
- Pre-warming: when a video goes viral, proactively push to CDN edge nodes
- Regional distribution: store popular video copies in S3 per region (to reduce cross-region CDN cost)
Key Design Decisions
| Decision | Choice | Reason |
|---|---|---|
| Upload | Chunked / resumable | Large files, unreliable networks |
| Transcoding | Async queue + workers | CPU-intensive; don't block upload response |
| Streaming | HLS/DASH with ABR | Industry standard; works on all devices |
| CDN | Multi-CDN (CloudFront + Akamai) | Resilience; optimize per region |
| Video storage | S3-compatible object store | Cheap, durable, globally replicated |
| Metadata | PostgreSQL | Rich queries; video/user/comment data |
| View count | Redis (approx) + batch to DB | High write rate; approximate count acceptable |
Security Considerations
| Threat | Mitigation |
|---|---|
| Malicious video upload (malware in container) | Virus scan raw uploads before transcoding; sandbox transcoding workers |
| Copyright infringement | Content ID fingerprinting (audio/video hash) on upload; compare against known database |
| CSAM / illegal content | Hash matching (PhotoDNA); AI classifier; human review queue |
| DRM / piracy prevention | Widevine/FairPlay DRM; signed URLs with TTL for premium content |
| Upload abuse (bot uploads) | Account verification; CAPTCHA; upload quota; ML detection |
| Signed URLs for private videos | Pre-signed S3 URLs with short expiry; validate token on each segment request |
| Transcoding resource abuse | Resource quotas per user; queue priority; kill runaway jobs |
| View count manipulation | Deduplicate views (IP + user agent + time window); ML anomaly detection on view spikes |
📰Real incident — YouTube creator account takeovers (2021). Google's Threat Analysis Group reported a long-running campaign that sent YouTube creators fake sponsorship offers. The "software to review" was cookie-stealing malware that took the creators' session cookies, which attackers used to take over channels — bypassing passwords and MFA — and then renamed them to run cryptocurrency scams or sold them. Google responded with more anti-session-hijacking checks for high-value accounts.
🎯On the job. For upload pipelines, ask how transcoding is isolated: video and image decoders parse attacker-supplied files and have a long history of memory-corruption bugs. Run transcoders in sandboxed, short-lived workers without access to credentials or internal networks.
Interview Tips
async pipeline is the right design and the interviewer wants to hear it.
"we split the video into segments at multiple bitrates, client picks based on bandwidth" — shows you know how streaming actually works.
here — 700M simultaneous streams cannot be served from a single origin.
mention Redis INCR per video, batch-sync to DB periodically — a classic approximation trade-off.
mention these as security/legal requirements. Interviewers at Google/YouTube will care.