Collection-time Sampling

Collection-time sampling means keeping only a fraction of events at the moment of ingestion and discarding the rest, to hold volume and cost down.

This differs fundamentally from analysis-time sampling: in analysis sampling all the data is stored and only a subset is computed for report speed, whereas in collection-time sampling the event is never stored in the first place and can never be recovered. That irreversibility is what makes it dangerous: a rare but important event such as a payment error can be lost in the sample and you would never learn how often it happened. So if a tool samples at collection, you need to know its rate and whether report numbers are scaled back up to true magnitude before display, otherwise every decision rests on data that itself does not know how incomplete it is.