An average is attractive because it compresses a large amount of activity into one number. That convenience is also its weakness. A route serving most requests quickly and a smaller group very slowly can still produce a reassuring average.
Begin with the shape, not the headline
Percentiles help describe how response times are distributed. The median offers a sense of the typical request; upper percentiles bring slower experiences into view. Neither is inherently more truthful. Their value depends on the question, route and traffic represented.
Look for distance between the centre and the upper tail, then ask whether that distance is stable. A widening tail may reflect one costly path, one region, a downstream dependency or a changing mix of work.
Keep the request population visible
Before comparing periods, check whether they contain similar routes, volumes and user behaviours. A busier but simpler traffic mix can look faster even when nothing in the application changed. A launch may introduce heavier workflows that make a broad service measure look slower.
Move from pattern to test
A useful reading ends with a narrower question: which routes contribute most to the slow tail? Did the distribution shift after one release? Is the effect concentrated in a region or time window? These questions create a bridge from analytics to technical investigation.