When latency numbers mislead
A short field note on averages, percentiles, and the questions that keep performance talk honest in analysis reviews.
Average latency hides the users who wait longest. When stakeholders quote a single number, ask which percentile, which region, and which time window. If those answers are missing, the number is decoration.
In Ha Giang cohorts we practice rewriting a performance claim into a constraint statement: “p95 checkout under X ms during peak ICT evening hours” is analysis language; “fast enough” is not.
Bring one chart you distrust to the constraint walkthrough. We will name what it can prove—and what it cannot.