Computing PPRs exactly can require extremely expensive preprocessing (in both time and space). Dynamic PPR works in the following way: during preprocessing, we compute "sketchy" random walk fingerprints for a small fraction of nodes, chosen using query log statistics. At query time, a small "active" subgraph is identified, bordered bynodes with indexed fingerprints. These fingerprints are adaptively loaded to various resolutions to form approximate personalized Pagerank vectors (PPVs). PPVs at remaining active nodes are now computed iteratively.