Keynote Interview

on YouTube ↗53m

" summary

Netflix runs a loosely coupled, seven‑layer personalization platform supporting model development, feature engineering and Java microservice serving for hundreds of millions of members; teams grew 5x since 2015 when Netflix had ~60M users. Key deliverables include a unified fact store, a Java library to avoid training‑serving skew for feature encoders, algo commons and Boson for shared encoders, and the open source Polynote. Iterations lead to Python turnkey APIs for data scientists, media ML for asset generation, and heavier guardrails, AutoML and continuous A/B automation.

AI-written summary · about

" chapters (12)

1. 0:00Keynote interview introduction
2. 0:41Netflix team and platform overview
3. 8:30Opinionated vs flexible platform
4. 12:42Evolving platform infrastructure
5. 17:56Platform architecture walkthrough
6. 22:30Open source role in platform
7. 26:01Deprecating in-house solutions
8. 27:49Media ML challenges
9. 36:52Automating ML to production
10. 41:50Fact store and communication
11. 45:42Fact vs feature distinction
12. 49:14Future opportunities in MLops

" related talks

• 04mInterviewShiva Chaitanya · Scale By the Bay 2017
• 03mInterviewJulie Pitt · Scale By The Bay 2018
• 03mInterviewVille Tuulos · Scale By The Bay 2019
• 29mEvolving the ML Platform Organisation at Netflix: A Case StudyJulie Amundson · YOW! 2021
• 22mJustin Basilico, Research/ Engineering Manager, NetflixJustin Basilico · MLconf SF 2015
• 29mBuilding Netflix - from movie magic to ML realityNagib Shah, Luke Ballantine, Matt Byrne · AWS Summit Sydney 2025
• 41mRecommendations for Building Machine Learning SoftwareJustin Basilico · Data By the Bay 2016
• 29mFrom encoding to machine learning a story of managing media data at NetflixArsen Kostenko · Highload fwdays'19