Sampling

Sampling means computing a report from a portion of the data instead of all of it and then generalizing the result to the whole; a way to make a report faster and cheaper at the cost of some uncertainty.

Sampling in analytics has two entirely different faces that must not be mixed. The first is sampling in reports: when data volume is high, some tools build a report from, say, 10% of sessions and then multiply up; this speeds the report but brings large error on small slices (a small country, a low-traffic page) and makes the numbers unstable. So the presence of report sampling is a quality warning; the user must know the number seen is an estimate, not a count. The second face is sampling in collection: the tool deliberately records only a share of events to control cost and volume (recording a share of sessions in replay, say). The key difference: report sampling is applied to full data that exists, whereas collection sampling records less data from the start, and what was not recorded does not come back.