Reproduce “Learned Offline Query Planning via Bayesian Optimization” on OpenResearch

About this paper

Jeffrey Tao, Natalie Maus, Haydn Jones, Yimeng Zeng, Jacob R. Gardner, Ryan Marcus

BayesQO introduces an offline query optimization approach using Bayesian Optimization to efficiently find high-performing execution plans. It maps query plans to a continuous latent space via a VAE, employs a dynamic timeout strategy based on model uncertainty, and consistently finds plans that execute significantly faster than those from traditional or other learned optimizers.

Key insights

• Bayesian Optimization (BO) is well-suited for offline query optimization due to its ability to efficiently explore expensive black-box functions and balance exploration/exploitation within a resource budget.
• Mapping discrete query plans into a continuous latent space using a Variational Autoencoder (VAE) enables the application of continuous optimization techniques like BO to the traditionally discrete plan space.
• Dynamically setting query execution timeouts based on model uncertainty and treating timed-out queries as censored observations significantly improves optimization efficiency by avoiding excessive time on poor plans while allowing promising ones to complete.

Summary by alphaXiv.