White-Box Caching for Cloud Data Systems

Modern data systems separate compute, storage, and memory, making data movement more expensive than computation. This dissertation introduces white-box caching: caches that inspect the structure, content, and access granularity of the data they store. It presents LiquidCache, SqueezeCache, and Bf-Tree, three systems that improve cloud analytics and transactional workloads through adaptive representations, lineage-guided cache content, and variable-length cache pages. The work also contributes to Apache DataFusion, Arrow, and Parquet.