The Flying Pig Goku Recommendation System was one of the most important projects during my time at Alibaba.From 0 to 1, this system eventually supported thousands of recommended scenarios for Flying Pig, including the 20 + venue of Double 11 in 2018.Today, I would like to share the design ideas and practical experience of this project.

# Project Background

Before the Goku system, the recommended scenarios for flying pigs were all independent, and each scenario had its own recommendation logic and code.This leads to a lot of redundant development, high maintenance costs, and difficulty in uniform optimization.We need a general recommendation system that can quickly support new recommendation scenarios through configuration.

# System Design

The core design concepts of the Goku system are "generalization" and "configuration".We have designed a unified recommendation framework that includes:

• • • Recall Layer * *: Supports multiple recall strategies, including collaborative filtering, content recalls, popular recalls, and more
• • • Sorting Layer * *: A unified sorting model that supports multi-objective optimization
• • • Rearrangement layer * *: Supports business rules such as diversity, freshness, etc.
• • • Configure the system * *: Define recommended scenarios through configuration without writing code

# Technical Challenges

In building this system, we encountered a number of technical challenges:

# # 1. Performance optimization

The system needs to support high concurrency, reaching tens of thousands of QPS during Double 11. We use slow