Скачать или смотреть YouTube Recommendation Engine: Complete Meltdown Analysis

YouTube went down. 350,000 users opened the app to a completely blank homepage — no recommendations, no Shorts, no Up Next. But search worked. Direct links worked. Ads ran fine.
That asymmetry is a blueprint.
In this video, I reverse engineer YouTube's recommendation system using the outage itself as the diagnostic — breaking down the two-stage ML pipeline that gets you from 800 million videos to 20 relevant results in under 200 milliseconds, why the blast radius landed exactly where it did, and what every systems designer should take away from a blank screen.

What we cover:
→ The scale constraint that drives every architectural decision
→ Two-tower neural networks and candidate generation
→ How the ranking stage works (and why it's a separate service)
→ The three layers of signals feeding the pipeline
→ Outage forensics: what failed and why the fallback didn't catch it
→ Graceful degradation — the design principle YouTube skipped
This is exactly the kind of system you'll be asked to design in senior engineering interviews.

Resources:
ByteMonk Blog: https://blog.bytemonk.io/
System Design Course: https://academy.bytemonk.io/courses
LinkedIn:   / bytemonk  
Github: https://github.com/bytemonk-academy

Timestamps
0:00 — The outage that revealed everything
1:15 — Why recommendations are uniquely hard at YouTube's scale
2:38 — Stage 1: Candidate Generation (two-tower neural networks)
4:03 — Stage 2: Ranking (and why it's a separate service)
5:25 — The three signal layers feeding the pipeline
8:45 — Outage forensics: reverse engineering the blast radius
10:30 — Graceful degradation: the design lesson from a blank screen
11:50 — What this means for your system design interviews

   • System Design Interview Basics  
   • System Design Questions  
   • LLM  
   • Machine Learning Basics  
   • Microservices  
   • Emerging Tech  

#YouTubeAlgorithm #systemdesign #bytemonk