Abstract

OLTP systems have significantly evolved since the 2008 study that identified CPU-bound tasks, such as buffer pool and concurrency control, as primary bottlenecks. In this paper, we revisit these assumptions and demonstrate that communication overhead is now the dominant factor affecting OLTP performance. Using comprehensive whole-stack benchmarks on modern hardware, we analyze both stored procedure and client-side transaction models, revealing that messaging costs play a critical role in end-to-end performance. Additionally, we explore user-defined code isolation, an often-overlooked aspect that poses security risks, especially in multi-tenant database systems. We find that increased isolation can lead to higher communication costs, highlighting the need for new strategies to mitigate these overheads.