The growth in the volume and scale of internet applications that rely on database systems meant that early database systems, which typically operated on a single node, rapidly reached their scalability limits. The traditional solution to address this limitation was to rely on a shared-nothing approach, wherein the data was partitioned across nodes, and each node was solely responsible for its data. When transactions needed to access data located at multiple nodes, the systems relied on some form of coordination to process such transactions. While such systems provided the scalability to meet the evolving needs of internet applications, they often produced poor performance under high contention. Further, depending on the underlying concurrency control mechanism, a single multi-node transaction could slow down a large subset of single-node transactions in the system. Nonetheless, for over a decade, shared-nothing systems provided the highest scalability and throughput for global applications.

In parallel, due to advancements in cloud service architecture, the concept of disaggregating monolithic systems into microservices has been adopted by the database community in the form of the disaggregated-storage database architecture. The Aurora system, which pioneered the model, addressed the scalability bottleneck of a single-node system by separating the database node into compute and storage layers. It minimized the amount of work performed by the compute node during transaction execution. Furthermore, the ability to replicate the storage layer permitted the creation of potentially infinite read replicas. Therefore, the primary compute node handled all writes, and the multiple read replicas provided unbounded read scalability. While such systems increased the overall throughput of the systems, they suffered low write throughput and high latency for global applications.

This proposal aims to bridge the gap between the two architectures by adopting shared-nothing techniques to support multiple writers in a disaggregated-storage setting. The proposal addresses the requirement in three parts. Firstly, it modifies an existing disaggregated-storage system to support multiple writers using optimistic concurrency control techniques without relying on intrusive changes at the database layer. The changes in the system, however, result in a regression in the cross-region read performance. The second part of the proposal addresses this limitation without affecting the performance gains produced by multiple writers. Lastly, the proposal aims to generalize the techniques of optimistic control to support write scalability of disaggregated-storage, irrespective of the underlying concurrency control technique used by the database layer.