A deep reinforcement learning model with plan value network for join order selection

The existing optimisers and dynamic programming methods rely on the cardinality estimation and the cost model of the local database. The resulting join plan does not reflect the execution time, and the errors of cardinality estimation will also lead to the join plans with poor quality. We propose a new learning optimiser, called DVJ (Deep reinforcement learning with plan Value networks for Join order selection). Compared with the existing deep reinforcement learning method, DVJ has two improvements: (1) the plan value network is designed to improve the reward mechanism in the existing deep reinforcement learning method, and the join plan generated by DVJ can reflect the latency; (2) applying Deep Q-Network to reduce the optimisation time and increase the chance of finding the best connection plan. Extensive experiments are conducted and the experimental results demonstrate that DVJ outperforms traditional optimisers and the existing deep reinforcement learning methods.