How Alibaba Cloud’s Milvus Service Boosted E‑commerce Search Stability and Scalability

Customer Introduction

ShiHuo, founded in June 2012, provides professional online shopping decision guidance, delivering timely fashion, lifestyle, and discount information across major domestic and international malls, helping users discover and purchase the latest, most cost‑effective fashion items.

Business Challenges

1. Rapid product category growth causing performance and relevance issues – Matching ShiHuo’s products with multiple sales platforms required precise association. Initial full‑text search with manual review could not meet recall and accuracy needs for the expanding long‑tail catalog, prompting a shift to a combined vector‑search, full‑text, and manual workflow.

2. Ensuring cluster stability while balancing cost and availability – Self‑built Milvus clusters experienced CPU usage fluctuations (50%‑100%) under load, affecting overall stability as traffic grew.

3. Lightweight management and operations to reduce complexity – Scaling self‑managed clusters increased monitoring, alerting, resource scaling, and kernel upgrade efforts, demanding significant manpower.

Alibaba Cloud Solution

Alibaba Cloud’s Milvus‑based vector search service is a fully managed, 100% compatible product that adds comprehensive operational infrastructure, enabling one‑click cluster deployment.

1. Significantly improved stability – Optimized read/write strategies balanced data distribution, raising query performance and achieving roughly a 10% QPS increase. Under a 2K TPS write load, CPU utilization remained stable around 50% without noticeable spikes.

2. Enhanced availability and flexibility – Over 100 monitoring metrics (including CPU and memory) and customizable alert rules provide precise cluster health insight. Flexible resource scaling allows smooth expansion or contraction to match business demand.

3. Substantial reduction in management cost – Although the managed cluster costs about 30% more than a self‑built setup, it eliminates the need for over one‑third of the personnel previously required for operations and infrastructure maintenance.

Architecture diagram of ShiHuo’s workflow:

Training and Inference with Alibaba Cloud PAI‑SAE

ShiHuo also utilizes Alibaba Cloud PAI‑EAS for embedding model training and inference. PAI‑EAS offers a one‑stop model deployment platform supporting CPU and GPU, delivering high concurrency and low latency for online recommendation scenarios.

Business Value

The Milvus service’s stable performance, distributed scalability, and diverse vector retrieval capabilities empower ShiHuo to handle large‑scale similarity searches essential for e‑commerce, while reducing operational overhead and positioning the team for future AI‑driven enhancements.

Product Updates

On July 19, 2024, the service was upgraded from the EMR Serverless Milvus beta to the dedicated Vector Search Service Milvus version, featuring an independent console, integrated Zilliz high‑performance kernel, and enterprise‑grade features.