MSP: Learned Query Performance Prediction Using MetaInfo and Structure of Plans

摘要：Query performance prediction is important and challenging in database management systems. The traditional cost-based methods perform poorly predicting query performance due to inaccurate cost estimates. In recent years, research shows that learning-based query performance prediction without actual execution has outperformed traditional models. However, existing learning-based models still have limitations in feature encoding and model design. To address these limitations, we propose a method of query performance prediction based on the binary tree-structured model fully expressing the impact between plan tree nodes. We also present an efficient metadata encoding method, taking into account the data type and value distribution of the columns, which we call metaInfo. This encoding method can support various complex SQL queries on changing data. The experiments are conducted on realworld datasets, and the experimental results show that our approach outperforms the state-of-the-art method.