The document describes a methodology for conducting large-scale online experiments using quantile metrics. It outlines three key challenges: 1) ensuring experiments are statistically valid and scalable given hundreds of concurrent tests involving billions of data points, 2) existing solutions are either statistically valid or scalable but not both, and 3) protecting users from slow experiences by detecting degradation in page load time quantiles like P90. The proposed solution provides statistically valid quantile estimates while being 500x faster than traditional bootstrapping through an asymptotic distribution approximation, compressed data representations, and optimized pipeline design that involves co-partitioning and aggregating histogram summaries across partitions. Results demonstrated the ability to analyze over 300 experiments, 3000 metrics, and up to 500 million members