E-LQO: A Comprehensive Energy Evaluation Framework for Learned Query Optimization

Abstract

Learned Query Optimization (LQO) has emerged as a promising paradigm for improving database performance, with reported speedups of 2× to over 10× on standard benchmarks. Existing evaluations, however, remain largely latency-centric and leave the energy debt from data collection, model training, and online inference less visible. This paper presents E-LQO, the first comprehensive framework for evaluating energy consumption across the complete LQO lifecycle: data collection or preparation, model training, model inference, and plan execution. We formalize two domain-specific metrics— Energy Return on Investment (EROI) and Energy Payback Time (EPT)—to quantify both per-cycle efficiency and the amortization horizon of learned optimizers. Across seven LQO methods on fixed-scale JOB and on TPC-H/TPC-DS up to 100 GB, we find that lifecycle energy is governed less by whether an optimizer is learned than by how it obtains supervision. Execution-based systems buy plan quality through active exploration and carry a large upfront energy debt; passive/log-driven systems can repay much faster with reusable logs or labels, but this advantage depends on incremental-cost accounting and often comes with weaker latency gains on light workloads. We further show that total online inference energy, including candidate-plan generation, enables symmetric accounting, and that roughly 50% template coverage is the energy-optimal exploration region for the execution-based methods we study. E-LQO provides practitioners with actionable guidance for energy-conscious deployment decisions and exposes optimization opportunities for sustainable learned query optimization.