Performance analysis of long-lived transaction processing systems with rollbacks and aborts

Increasing the parallelism in transaction processing and maintaining data consistency appear to be two conflicting goals in designing Distributed Database Systems (DDBS). This problem becomes especially difficult if the DDBS is serving long-lived transactions (LLTs). Recently, a special case of LLTs, called sagas, has been introduced that addresses this problem. The DDBS with sagas provides high parallelism to transactions by allowing sagas to release their locks as early as possible. However, it is also subject to overhead due to efforts needed to restore data consistency in case of failures. In this paper, we first conduct a series of simulation studies to compare the performance of LLT systems with saga implementation (or saga systems) and the LLT systems without saga implementation (or nonsaga systems) in a faulty environment. The simulation studies show that the saga systems outperform their nonsaga counterparts under most of conditions including the heavy failure cases. We thus propose an analytical queuing model to further investigate the performance behavior of the saga systems. The motivation of the development of this analytical model is twofold. It assists us to further study quantitatively the performance penalty of the saga implementation due to the failure recovery overhead. Furthermore, the analytical solution can be used by system administrators to fine tune the performance of the saga system. This analytical model captures the primary aspects of the saga system, namely, data locking, resource contention, and failure recovery. Due to the complicated nature of the analytical modeling, we solve the model approximately for various performance metrics using decomposition methods, and validate the accuracy of the analytical results via simulations.