ELENA: AN Explainability-Aided Online Query Optimization Framework

In recent years, learned query optimizers have achieved remarkable progress in reducing overall execution latency by leveraging machine learning methods and large language models. However, in online scenarios, the continuous accumulation of training samples results in substantial retraining overhead. Though some studies try to reduce the training scale through sample selection strategies, such as choosing the most recent samples or using another grading model to decide their reservation, these approaches cannot guarantee the quality of the selected samples because the selection process lacks interpretability. Therefore, achieving high query performance with low training costs in online scenarios remains a critical challenge. To address this limitation, we propose an Explainability-aided onLine quEry optimizatioN frAmework, Elena, designed to achieve high query performance while minimizing retraining overhead. Elena designs two interpretable modules: (i) Dynamic sample selector which utilizes influence score to retain the highquality samples from the historical samples and applies a tailored similarity measurement to capture valued and diverse new samples, while guaranteeing the bounded training set size by theoretical proof; (ii) Adaptive feature selector which automatically identifies the features most critical to model performance, guiding the training and inference towards active features to improve the performance. Experiments across multiple public benchmarks demonstrate that Elena significantly reduces the retraining cost of learned optimizers by up to 67% while maintaining performance, resulting in up to 18% reduction in plan execution time. These results highlight that Elena can achieve novel performance under an extremely limited training set.