Each node stores a routing table of only about √n log n entries, not the full
n. Routing uses only those tables and the target's label, yet the path taken is never more than
3× the true shortest path. Pick a source and a target below and watch the route build hop by hop.
How it works. A random set A of landmarks is sampled, each node with probability 1/√n.
Every node knows its nearest landmark (its centroid) and keeps a routing-table entry for every landmark and for every
node in its cluster (the nodes closer to it than to their own landmark). A target's label carries its centroid
and the first hop out of it. If the target is a landmark or sits in the source's cluster, the route goes straight there;
otherwise it detours through the target's landmark. That single detour is what bounds the stretch at 3.