Adaptive database's performance tuning based on reinforcement learning

Database (DB) performance tuning is a difficult task that requires a vast amount of skill, experience and efforts in tweaking a DB for optimum results. With the hundreds of parameters to be considered under the diverse application configurations, business logic and software technology, getting a true global optimum setting is difficult for a DB administrator. We propose a novel approach based on Reinforcement Learning to tune a DB adaptively with minimum risk to the production setup. It results in a new set of parameters tailored to the production DB. Empirical results show that there is a significant gain in performance for the DB in its overall efficiency while reducing the IO overheads, based on a set of key performance statistics collected before and after the optimization process