Flash Sale Engineering

on YouTube ↗19m

" summary

Engineered a queuing layer in the load balancer using Nginx + OpenResty Lua, a per-LB leaky bucket that only admits a calculated share of global checkout throughput (example capacity ~16), and a timestamp cookie plus a PID lag eliminator that shifts priority so early arrivals get served first. Back pressure is served as a cached throttle page with background polling, and end-to-end load tests cut p95 queue time from roughly the sale duration (~30 minutes) to a few minutes.

AI-written summary · about

" chapters (7)

1. 0:00Start
2. 1:28Flash Sales
3. 2:26Urban Outfitters
4. 3:02Notable Flash Sellers on Shopify
5. 4:12Overview of How Shopify Works
6. 9:15Feedback System
7. 16:31Lag Indicator

" related talks

• 42mA Commerce-centric Approach to Queuing Fairly at High ThroughputLogan Martel · Strange Loop 2022
• 41mLoad Balancing RoundupAlan Kasindorf · Velocity 2009
• 42mPredictive Load-Balancing: Unfair but Faster & more RobustSteve Gury · Strange Loop 2017
• 47mFrom Legacy to High-Load: Evolution of a Web ApplicationYevgen Lysenko · Highload fwdays'18
• 38mA Practical Look at Performance TheoryKavya Joshi · Strange Loop 2018
• 50mBuilding and Operating Online Shops under Extreme LoadPhilipp v. Criegern · code.talks 2017
• 42mStop Rate Limiting! Capacity Management Done RightJon Moore · Strange Loop 2017
• 19mCome Hell or Still Water: Alleviating Tail Latency in Cloud Block StoreYifan Shen, Haoran Zhang, Ennan Zhai, Chaolei Hu, Fengyuan Ren, Erci Xu, Lingjun Zhu, Kun Qian, Xue Li, Yuesheng Gu · NSDI '26