A release marker invites a simple comparison: before and after. But the nearest windows are not always comparable. Time of day, weekday, campaign traffic and background workload may matter as much as the deployment.
Choose windows that represent similar work
Start with traffic shape. Compare route mix, request volume and the proportion of important user journeys. Where possible, use matching weekdays or operational periods and document unusual events.
Allow for warm-up and delayed effects
Some changes appear immediately; others emerge as caches turn over, queues build or a new feature gains adoption. Excluding a short transition period can avoid confusing deployment activity with steady behaviour, while a longer follow-up window can reveal delayed movement.
Describe confidence, not just direction
A small change across few requests should not carry the same weight as a sustained shift across a large, comparable sample. State sample size, variability and known changes in traffic. The aim is a defensible interpretation, not a dramatic verdict.