In this paper, we revisit the question of what a scalable OLTP design should look like. The industry has mainly settled on a shared-log design, where the log of updates is applied to secondaries to make them available for reads. An alternative is a shared-caching system where every compute node can access all the storage and cache all the data that a transaction reads and writes. The latter provides exciting opportunities for the cloud, such as scalability for skewed reads while avoiding user-defined partitioning and two-phase commit. Moreover, we discuss lessons learned from building a new cloud OLTP DBMS based on a shared-caching design and present research opportunities for the community.