Stratified Sampling
Splitting or sampling so each group appears in the same proportion as it does in the full dataset.
Usually stratified on the target, so every fold and every split carries the real class balance. It's the default for classification, and it stops mattering only when every class is abundant.
The reason is variance rather than fairness. Take 1,000 rows at 2% positives and a random 20% test split expects 4 positives with a standard deviation of 1.77, so across ten seeds you might see anywhere from 2 to 5. At 2 positives, one classification moves recall by 50 points. Stratifying gives exactly 4 every time.
It also applies to groups other than the target — region, source, device — whenever a fold missing one of them would produce a score you can't interpret.