Design and Optimization of Taobao's Large‑Scale Flash Sale (秒杀) System

Taobao's flash‑sale (秒杀) system originated from a timed product‑listing feature that caused sudden traffic spikes; to protect the main product‑detail service, a dedicated high‑concurrency system was created.

Key data : During the 2013 Xiaomi flash‑sale, front‑end QPS exceeded 600k, backend cache clusters peaked at 20 million requests per second, and the highest order‑processing TPS reached 1,500 /s.

Hotspot isolation is achieved at three levels: business isolation (separate marketing activity registration), system isolation (dedicated deployment groups and domain names), and data isolation (exclusive cache clusters or MySQL instances for hot data).

Static‑dynamic separation caches 90 % of page data on the client or CDN, leaving only the "refresh‑抢宝" button to request minimal dynamic data, which improves performance by over three times.

Time‑slice throttling adds a quiz before purchase to stretch the order burst from 1 s to 2‑10 s, reducing the impact of automated bots and smoothing the load.

Real‑time hotspot discovery collects hotspot keys from middleware (Tengine, Tair, HSF) within 3 s, publishes them via a hotspot service, and allows downstream systems to apply protective measures.

Key technical optimizations include:

Using raw Servlets and output streams instead of heavyweight MVC frameworks.

Local in‑process caches for static product data and short‑lived caches for inventory.

Layered data validation (static‑dynamic split, client‑side cache, web‑side cache, write‑side checks, DB consistency).

Queue‑based write handling at both application and MySQL InnoDB layers, with Alibaba‑contributed patches for row‑level concurrency control and fast commit/rollback.

Overall principles for large‑scale traffic systems are isolation, dynamic/static separation, and layered validation, applied across the entire request chain to ensure stability and performance.