Cloud-native data warehouses and other DBMS decouple storage from compute: an elastic compute layer accesses data stored remotely in block-oriented columnar format. To deliver interactive performance given the high-latency and low-bandwidth to remote storage, most cloud-native DBMS cache data at the compute node. In this paper, Durner et al. propose a smart cache storage system, Crystal, which sits between the database and/or execution engine (e.g. Spark, Greenplum) and raw storage. Crystal includes a CMS that runs in the compute node and can interact with remote storage; and DBMS-specific clients, which implement the data source API with push-down predicates. Crystal, which identifies which regions to transform and cache locally, achieves lower query latencies and more efficient use of bandwidth between compute and storage.