Field note
Choosing an observation window for a workload study
An observation window should cover at least one full business cycle for the application—often a weekday pattern plus a weekend contrast, or a billing period if usage is invoice-driven.
If you already know a campaign or release will reshape traffic, schedule the window to include a quiet baseline and the event itself. Studying only the event without baseline context makes every spike look unprecedented.
Access constraints matter. Prefer read-only telemetry export and agreed sampling over intrusive production experiments. If synthetic probes are needed, keep them labeled so they never contaminate customer-facing statistics.
Close the window with a freeze on interpretive debate: circulate draft findings, collect corrections on facts, then lock the narrative for the walkthrough. Endless re-slicing of the same metrics rarely improves decisions.