From encoding to machine learning a story of managing media data at Netflix
summary
Netflix builds a media intelligence platform that turns ingestion into a managed, versioned dataset so models actually work in production: initial duct‑taped JSON in Cassandra fails, so a snapshot view plus a validated common format and an API enforce schema‑on-read, track who changed what, and produce labeled datasets for QC, thumbnails, dead‑pixel and subtitle‑collision checks. It treats ML as a small component and gives concrete operational controls (HBase/Cassandra backing, workflow tips like Dynamo for Conductor) and it works.
AI-written summary · about
chapters (35)
- 0:00Intro
- 0:59Who is Arsen
- 1:44What is Media Intelligence
- 2:59Dead pixels
- 3:45Number of movies
- 4:38Movie thumbnails
- 4:49Banners
- 5:25Object on other side
- 5:38Image like this
- 5:57Another interesting problem arrives
- 6:22How does industry solve this
- 6:50Average breakdown
- 7:43Drakkar Pal
- 9:06adversarial attack
- 10:33data infrastructure
- 11:45itzik routing
- 12:21machine learning
- 12:57Anderson Horowitz
- 13:17Focusing on the data
- 14:47What changed
- 15:05What do we want
- 15:26Back to Netflix
- 15:50Time dimension
- 17:05Space dimension
- 17:38Data structure of programming
- 17:50First idea
- 18:17The problem
- 19:12Different teams
- 19:45Common Data Structure
- 21:09The Snapshot
- 21:44Two Things
- 22:03Know Your Problem
- 22:33Make a Picture
- 22:57Gradient Descent
- 23:12Recap
