How Should Applications Choose Optimal Isolation Levels?

Abstract

Modern relational databases provide multiple isolation levels, yet achieving full serializability often incurs significant performance overhead. Recent research shows that, by restructuring specific query patterns, workloads can preserve serializability even under weaker isolation levels. However, existing approaches often overlook the trade-off between the performance gains of lower isolation levels and the additional coordination or enforcement mechanisms required to maintain correctness. This can result in suboptimal isolation choices that fail to maximize overall system performance. In this talk, I will present middleware-based techniques that enable serializable scheduling through self-adaptive isolation level selection. By dynamically balancing correctness and performance, these techniques allow application developers to focus on business logic while the middleware transparently ensures consistency and efficiency.