Reasoning about data flows across large codebases — to enforce security policies, map sensitive data for privacy and compliance, and support data-governance requirements — demands automated analysis that is simultaneously scalable, precise, and sound. This talk presents our experience building and deploying a static taint analysis engine at Amazon that addresses these competing demands through compositional, bottom-up data-flow summarization: analyzing individual program components in isolation and combining the results modularly. We discuss the key design choices that make this practical at scale, including strategies for modeling un-analyzable code, generating actionable data-flow traces, and providing analysis diagnostics that help both developers and analysis engineers understand and trust the results. We also reflect on where the approach succeeds, where it falls short, and what extensions we are pursuing.