Online A/B experiments require timely detection of harmful changes while controlling false alarms. We introduce YEAST, a simple sequential test for continuous monitoring of count and real-valued metrics over a prespecified event horizon. It compares the cumulative difference between experiment groups with a constant boundary determined by an estimate of the variance of this difference at the end of monitoring. Under the strong null of no effect on the event-generation process, we establish asymptotic Type-I error control through a maximal inequality and a central limit theorem. The required conditions allow some forms of repeated-user dependence. We evaluate YEAST using synthetic simulations and a semi-synthetic study of fixed-duration monitoring and observe power gains over the tested martingale-based methods.

翻译：暂无翻译