Field notes
Choosing a baseline window before you judge a release
Deployment impact analytics lives or dies on the baseline. Compare a Tuesday mid-morning release to a quiet Sunday and almost any change looks dramatic.
Match the traffic shape
Prefer a baseline from the same weekday and roughly the same hours as the post-deploy window. If your product peaks at lunch in Korea, a midnight baseline will mislead you about latency and error rates.
Account for prior deploys
If another release landed inside your intended baseline, shorten the window or skip that day. Overlapping deploys muddy attribution: you cannot honestly assign a latency bump to the newer change when the older one is still settling.
Document exclusions
Mark marketing campaigns, batch jobs, and regional holidays. In Gwangju and nationwide Korean holidays, traffic often dips — that dip is not evidence your release improved capacity.
Practical rule
We usually recommend one clean business day of comparable traffic before the deploy, plus a same-length window after. When that is impossible, say so in the findings brief instead of forcing a neat chart.