Cassandra has no leader. Every node is equal, and the ring decides who owns what. Each node is assigned a set of tokens — positions on the ring — and by default there are 256 of them.
Adding a node claims tokens from many existing nodes at once. They all stream a little data, and nobody streams all of it.
Losing a node is equally undramatic. Its 256 ranges are already replicated elsewhere, so the whole cluster absorbs the load rather than one unlucky neighbour.
It also makes heterogeneous clusters work. Give a bigger machine more tokens — 512 instead of 256 — and it takes a proportionally bigger share. The machines no longer have to be identical to be treated fairly.
There is a cost. More tokens mean more metadata gossiping between nodes, and repair operations get slower. Cassandra later reduced the default for large clusters, which tells you the number was a trade-off all along rather than a constant.
The elegant abstraction and the operational reality are rarely the same number.