Four Hops Out

A friend recommendation feature walks four hops out from a user. In Postgres the recursive CTE over a 60 billion row edge table takes minutes. What does a native graph engine change about the traversal itself?

Each vertex stores direct references to its neighbors, so a hop costs work proportional to that vertex's degree instead of an index probe into the global edge table

It holds the entire graph in RAM, and that residency is the only mechanism by which any traversal can beat a relational join running over the same edge data on disk

It rewrites the recursion into a single hash join over the whole edge table, so the depth requested by the traversal stops mattering to what the query costs

It materializes every path up to a configured depth as edges are written, so a traversal at query time becomes a plain lookup by the pair of endpoint keys