Building a Personalized Product Recommender System with Vipshop and Milvus

This post details Vipshop's implementation of a personalized product recommender system using Milvus. It describes a read/write separation architecture, the use of Redis for managing collection aliases to enable seamless data updates (effectively near real-time updates by switching entire data collections), batch writing data to multiple partitions, index building and pre-loading for performance, and a data recall process that achieves an average latency of 30ms. It also highlights lessons learned regarding connection pooling for the Java client, index warm-up, and parameter tuning (nlist, nprobe).