GNN serving at scale

Design online serving for a graph neural network that powers 'people/items you may know' on a social graph with 1B nodes and 100B edges. The GNN's prediction for a node depends on aggregating its multi-hop neighborhood, so a naive online forward pass would have to fetch and process an exploding number of neighbors per request — infeasible within a serving latency budget. Design the system so GNN-powered recommendations are served at scale and low latency, and explain how you avoid the neighbor-explosion problem online.