21 April 2026
Sampling without losing the plot
Warehouses eventually refuse to scan everything before lunch. Sampling arrives dressed as pragmatism. The danger is not the sample. It is the silence around who was eligible and who was too rare to appear.
We treat a sample as a map inset, not as the territory. First write the population: first-valued users in a window, excluding staff, excluding a known broken build. Then draw the sample with a seed you record. If you cannot replay the draw, you do not have a method; you have a screenshot of whoever loaded first.
Rare paths vanish under uniform sampling. The people who hit a forgotten deep link, or a regional paywall, or a voice-only save, will not show up in a 2% draw, and the readout will declare the path unimportant. Stratify on the surfaces you already fear. If you do not know which surfaces to fear, you are not ready to sample; you are ready for Event Grammar.
Always keep a count of the unsampled remainder. A landscape that only describes the convenient 50,000 is a brochure. In the Studio we ask for two numbers on every sampled chart: the draw size and the population it claims to represent. Missing the second number is treated as an incomplete assignment, not a style choice.
Tool defaults are not innocent. Some product analytics platforms sample in the UI while the export is complete, or the reverse. Know which layer you are in. A disagreement between a tiled number and a warehouse number is often two different samples wearing the same event name.
Sampling is allowed. Forgetting the people outside the inset is not. If that sentence feels heavy, it is doing the work a dashboard legend was supposed to do.