Data Sampling

Your analytics pipeline processes 500 million events per day. Computing exact metrics (unique visitors, average session duration, conversion rate) on every single event costs $3,000/month in compute. But your dashboard only needs to be accurate within 2%.

A 10% random sample of 50 million events gives you statistically accurate results (within 1-2% error) at one-tenth the cost. For most analytics use cases, "approximately 2.3 million unique visitors" is just as useful as "exactly 2,317,489 unique visitors."

Data sampling is the practice of analyzing a representative subset of your data instead of the entire dataset. It works when processing the full dataset is too slow, too expensive, or simply unnecessary for the required accuracy.

Google does not analyze every search query for trend detection. They sample. Facebook does not compute exact friend-of-friend counts for every user. They estimate from samples. Spotify's Wrapped feature processes listening history for 600M+ users, and even they sample intermediate calculations to keep compute costs manageable. Sampling is a pragmatic engineering trade-off: less accuracy, dramatically less cost.

Sample size matters. Too small and your results are unreliable. Too large and you have not saved much. For most metrics with normal distributions, 10,000 samples gives you under 1% error. For rare events (fraud, errors occurring at 0.001%), you need larger samples or targeted sampling that oversamples the rare category.