You can now choose percentile analysis on a metric and use it in experiments, measuring P50 through P99 instead of only the mean. For latency, load time, or any long-tail measure, an average smooths over the slow experiences users actually notice, and a change that improves the median can quietly worsen the 95th percentile without moving the headline number. Choose the percentile when you define the metric: P50 and P75 read general trends, while P95 and P99 surface outliers, such as an endpoint that behaves well until it meets your largest customer. Results show mean and percentile metrics side by side so they're easy to compare.

Percentile metrics work with hosted and warehouse native experiments on both Bayesian and frequentist analysis, though frequentist experiments must use sequential testing.

Read Percentile analysis (https://launchdarkly.com/docs/home/metrics/components#percentile-analysis) for more info.