B-trees keep their balance by refusing to let any node overflow. This diagram shows insertion as a two-phase operation: a descent and a cleanup. The descent is a sequence of comparisons — at every internal node the database picks the child whose key range brackets the new key, and it always ends at a leaf. The cleanup is where the structure is preserved: an overflowing node splits in half, promoting its middle key to the parent, and the promotion can cascade all the way to the root. A root split is the only time the tree grows, and it adds exactly one level, so every leaf stays at identical depth.
That uniform depth is the entire reason databases use B-trees: lookup, insert, and delete are all O(log n) with the log based on the node's fanout, and the tree never needs the rebalancing rotations that avl-style structures require. Each stage here is a page-sized unit of work, which is exactly how the database thinks about it — one page at a time.