Some data is too large to read end-to-end. Sublinear algorithms answer questions about it anyway, and tell you exactly how much to trust the answer.
Sketches and streams
A sketch is a compressed representation that keeps just enough structure to answer a specific question: Morris counters that count to a billion in a handful of bits, count-min sketches that track frequencies across a stream, locality-sensitive hashing that finds near-duplicates without pairwise comparison. You choose the accuracy and the confidence; the algorithm chooses how little it can get away with reading.
Why it pays
This is the mathematics under our data work: profiling and quality-scoring estates too large to scan is exactly the regime where randomized approximation with guarantees beats exhaustive computation. We build these primitives in-house (approximation schemes with tunable ε and δ) and they end up inside NRV.