back

How Do Learned Cost Models Improve Your Database Performance?

How Good Are Learned Cost Models Really? Insights from Query Optimization Tasks

2025/08/07

Published in SIGMOD 2025

Authors: Roman Heinrich, Manisha Luthra Agnihotri, Johannes Wehrstein, Carsten Binnig

Cost models are the cornerstone of query optimization, guiding database systems to select highly performant execution plans. Recently, Learned Cost Models (LCMs) demonstrated high accuracy in predicting query costs. However, a crucial question remains open: How does query optimization profit from this accuracy provided by LCMs? This is what we systematically explore in our brand-new paper “How Good Are Learned Cost Models Really? Insights from Query Optimization Tasks”, that appeared at SIGMOD 2025..

Learn More

Learn more in our LinkedIn article.

The paper can be found here.

Example Query Nr. 33 from JOB-Light with different join orders: Postgres picks the optimal plan and outperforms all LCMs for most of the metrics.
Example Query Nr. 33 from JOB-Light with different join orders: Postgres picks the optimal plan and outperforms all LCMs for most of the metrics.