Bringing modular concurrency control to the next level

Database users face a tension between ease-of-programming and high performance: ACID transactions can greatly simplify the programming effort of database applications by providing four useful properties—atomicity, consistency, isolation, and durability, but enforcing these properties can degrade performance.

This dissertation eases this tension by improving the performance of ACID transactions for scenarios where data contention is the bottleneck. The approach that we take is federating concurrency control (CC) mechanisms. It is based on the observation that any single CC mechanism is bound to make trade-offs that cause it to perform well in some cases but poorly in others. A federation opens the opportunity of applying each mechanism only to the set of transactions or workloads where it shines, while maintaining isolation.

In particular, this work builds upon Modular Concurrency Control (MCC), a recent technique that federates CCs by partitioning transactions into groups, and by applying different CC mechanisms in each group.

This dissertation addresses two critical shortcomings in the current embodiment of MCC. First, cross-group data conflicts are handled with a single, unoptimized CC mechanism that can significantly limit performance. Second, configuring MCC is a complex task, which runs counter to MCC’s purpose: to improve performance without sacrificing ease-of-programming.

To address these problems, this dissertation presents Tebaldi, a new transactional database that brings Modular Concurrency Control to the next level, both figuratively and literally. Tebaldi introduces a new, hierarchical model to MCC that partitions transactions recursively to compose CC mechanisms in a multi-level tree. This model increases flexibility in federating CC mechanisms, which is the key to realizing the performance potential of federation. Tebaldi reduces configuration complexity by managing the MCC federation automatically: it can detect performance issues in the current workload in real-time, and automatically adjusts its configuration to improve its performance.