Reducing Workload On Real-Time Database Using Machine Learning Based Transaction Scheduling

Abstract

Reading and writing to relational databases requires accessing multiple tables for constraint & quality checks. In order to perform these checks, databases use transaction management, wherein index-based checking & validation is done, and data is committed to the database only when these checks are satisfied. In case of any validation violations, databases need to either fall back to previous data state, or activate violation rule engine and resolve the underlying conflicts. Performing these tasks for limited size databases doesn’t compromise on system performance, but as database sizes increase, the number of checks increase exponentially, thereby reducing database system performance. In order to reduce the effect of database size on transaction scheduling performance, this work proposes a genetic algorithm inspired algorithm, which takes into consideration multiple performance parameters in order to optimize transaction performance. The underlying system is deployed on multiple relational databases, and a performance improvement of 10% in terms of scheduled transaction execution delay is observed. This performance is compared with recently proposed state-of-the-art systems, and it is observed that the proposed model is able to reduce execution delay by 5% across multiple implementations.