Eugene Yan's talk for the MLOps Community walks through the core architecture of recommendation and search systems. The emphasis is on a two-stage pipeline: first, candidate retrieval narrows a vast corpus to a manageable set of promising items; second, a ranking model orders those candidates by relevance or predicted engagement.
Yan grounds the discussion in industry examples, showing how companies make concrete trade-offs. For instance, retrieval must be fast and high-recall, while ranking can be more computationally heavy because it only processes a shortlist. This separation lets teams optimise each stage independently.
The talk is less about novel algorithms and more about engineering practice—how to think about system constraints, data flows, and serving infrastructure. For practitioners, it serves as a useful checklist when designing or debugging a recommendation or search stack.