Limitations of model-based database knob tuning : Problems with Bayesian Optimization and solutions using Hyperband

Automatic database tuning is an emergent research area which applies hyper-parameter optimization techniques for tuning settings in database management systems. The best performing current approaches use Bayesian optimization (BO) with a Matérn kernel even though some research suggests that a non-model approach could be better. This thesis shows that BO is unable to accurately model the underlying black-box functions of database tuning and that the novel database tuning approach Hyperband is more general. Both BO and Hyperband were evaluated on two optimization tests for different workloads showing that Hyperband is a broadly efficient approach with short search times while the ef-fectiveness of BO depends inversely on cost of the query being optimized.