Recommendation feed generation and ranking

Design the full candidate-generation to re-ranking pipeline for a recommendation feed over a corpus of 1B items for 300M users, returning a feed of 50 items in under 150ms. You can't score 1B items per request, so describe the funnel: how each stage narrows the corpus, what model/index each stage uses, and how you keep the stages from fighting each other (e.g. the re-ranker can only re-rank what generation surfaced). Where does business logic / diversity / freshness get injected?