Related data has gravity.

Mark one field as the anchor. Every record that shares its value — a customer with their credits, payments and history — is written into the same block on disk, so pulling that customer's whole graph becomes one sequential read instead of a fan-out of JOINs.

Underneath it's a Z-order curve: a record's x, y and z collapse into a single sort key, so things that mean the same end up ordered as neighbours, on the same page.