The Curse of P90: An Elegant Way to Overcome It Without Magic

摘要

Over the decades of development of methodologies and metrics for IT capacity planning and performance analysis, percentile terminology has become the lingua franca of the field. It makes sense: percentiles are easy to interpret, not sensitive to outliers, and directly usable for approximating the distribution of the variable being measured for stochastic simulations. However, depending on which percentile is used, we can miss important information, like multimodality of the metric's distribution. Another, less obvious, downside of relying on percentiles comes into play when we size infrastructure for a high percentile of demand (e.g., p90). Given that it takes time to order, manufacture, receive, and install infrastructure, this means that we need to answer the statistically nontrivial question, ''what will this percentile of demand be a few years from now?'" This paper discusses the issues that arise in answering it and proposes an elegant way of resolving them.