Designing Tuners with Explicit Generalizability Objectives

Develop automated database tuning frameworks that jointly optimize for performance and generalizability across dataset sizes and environments to avoid configurations that overfit to small instances and fail to transfer effectively to larger deployments.

Background

In dataset growth experiments, the authors find that certain performant plans at small scale (SF1) fail to transfer to larger scale (SF20), highlighting the need to incorporate generalizability objectives in tuner design.

They explicitly defer building tuners with such objectives, underscoring an open problem of balancing performance and robustness across changing data volumes.

References

We defer building tuners with an explicit generalizability objective, in addition to a performance objective, for future work.

— This is Going to Sound Crazy, But What If We Used Large Language Models to Boost Automatic Database Tuning Algorithms By Leveraging Prior History? We Will Find Better Configurations More Quickly Than Retraining From Scratch! (2510.17748 - Zhang et al., 20 Oct 2025) in Section 6.6 (Evaluation: Dataset Growth)