# This is how Netflix recommends fresh content in the homepage.

## Метаданные

- **Канал:** Gaurav Sen
- **YouTube:** https://www.youtube.com/watch?v=hoXZ3SxFZrQ
- **Дата:** 02.04.2026
- **Длительность:** 1:21
- **Просмотры:** 49,672

## Описание

System Design Course at InterviewReady: https://interviewready.io

## Содержание

### [0:00](https://www.youtube.com/watch?v=hoXZ3SxFZrQ) Segment 1 (00:00 - 01:00)

This is how Netflix reduced its average latency for recommendations by 12%. Netflix streams 16 billion hours of content every month. The home page is what drives this. It recommends fresh content to every user. The ranker service does this by comparing new titles against what you've already watched. It uses dot products between vectors to measure how similar two titles are. This process would consume 7 and 1/2% of CPU on every Netflix recommendation node. Netflix was doing this with basic Java multiplication. Every dimension of the vector multiplied and added one at a time. To speed this up, one idea was to try the BLAST library, which is basic linear algebra subprograms. These are optimized libraries for mathematical competition. But every call crosses the Java native interface boundary. That means copying data back and forth between Java and native code. The overhead was so bad, they abandoned it. So, Netflix tried the Java vector API instead. This uses SIMD, which is single instruction multiple data. Instead of multiplying one dimension at a time, SIMD multiplies four or eight at once. The result was that Netflix saved 7% across their CPU clusters, and there was a 12% drop in average latency. At Netflix's scale, serving petabytes a day, that's a massive saving. Thanks for watching. Cheers.

---
*Источник: https://ekstraktznaniy.ru/video/49977*