System 12.12 — Netflix
In one line
YouTube's streaming + the recommendation/open-connect story: regional caching boxes and ML-driven home.
Think of it like this
YouTube's harder cousin: instead of user uploads, it's about serving pre-processed shows to tens of millions of people watching at the exact same peak hour (Friday night) without buffering.
Key ideas
- 01
Two halves: content pipeline (ingest + transcode, like YouTube) and the personalized home feed (recommendation service).
- 02
Open Connect: Netflix's own CDN boxes inside ISPs — serving 95% exchange of traffic with cache hits.
- 03
Recommendation: candidate generation (collab + content vectors) → ranking model → personalized row feeds per profile.
- 04
Watch history feeds the model async (events via Kafka to feature pipelines — offline batch + online serving).
- 05
Consistency: catalogs are eventual; a new title appears on CDN caches regionally within hours.
- 06
Billing/preferences: classic CRUD service + entitlements (who can watch what per plan).
- 07
Interview angle: most of this is 'media pipeline + personalization'; presentation matters more than novelty.
- 08
Scale numbers to carry: 200M+ subs, petabytes/day of ingestion, CDN-first delivery.
Code & diagrams
The upload pipeline looks like YouTube's; the READ side is the whole interview — millions watching the same peak hour.
Explain without notes
Why does Netflix put hardware inside ISPs instead of renting big CDNs — what does it buy them operationally?
Practice
Design the recommendation data flow: events → features → model → ranked rows → cache → UI.
Trade-offs
- ↔
Self-built CDN = huge ops cost + total control of the last mile; the classic build-vs-buy at scale.
Run it in production
You've designed it. Now build, operate, and break the same idea hands-on in the DevOps courses:
Completion checklist
I can structure Netflix = content pipeline + personalization + Open Connect and defend each.