Data Sampling

Data sampling means analyzing a representative subset of the data instead of the whole, and generalizing the result — a method for speed and economy, at the cost of some uncertainty.

Sampling is not inherently bad; if the sample is drawn correctly and is large enough, its result is very close to the true number while saving substantial resources. The problem is when sampling becomes hidden and forced: some tools, as data volume rises, switch to a sample without telling the user, and the reported number is no longer exact without the user knowing. An important distinction is between two uses of sampling: sampling in computation (such as estimating unique users with a statistical algorithm), a deliberate technical choice, and sampling in storage or display (dropping some events), which directly affects trust in the number. The healthy rule is: sampling is fine so long as it is transparent and the user knows whether the number they see is an estimate or a full count.