Chunshan Zhao’s paper on learned cardinality estimation, titled “How Good Are Learned Cardinality Estimators, Really? A Comprehensive Experimental Study,” has been accepted by SIGMOD 2027.

Cardinality estimation is a fundamental component of database query optimization, as the quality of cardinality estimates directly affects the execution plans selected by query optimizers. In recent years, learned cardinality estimation (LCE) has shown substantial potential by leveraging machine learning to capture complex data distributions and query patterns, improving estimation accuracy and, in some cases, end-to-end query performance. However, the practical effectiveness of learned cardinality estimators in real-world database systems, particularly across multiple deployment-related dimensions, has yet to be systematically evaluated. To address this gap, this paper presents a comprehensive experimental study of ten representative LCE methods on three real-world datasets. Beyond overall effectiveness in terms of estimation accuracy, plan quality, and end-to-end query latency, the study systematically investigates five practical dimensions: model construction and storage overhead, training data efficiency, generalization under query workload drift, scalability, and adaptability to dynamic data updates. It also develops a unified taxonomy and analysis framework covering both query-driven and data-driven approaches. The experimental results show that neither query-driven nor data-driven methods clearly dominate, and no single learned estimator consistently performs best across all evaluation dimensions. Moreover, estimation accuracy alone provides an incomplete picture of practical effectiveness: plan-ranking ability, end-to-end performance, deployment overhead, robustness across workloads, and adaptation to data updates are also critical. While learned estimators can significantly improve accuracy and query performance under favorable conditions, consistently outperforming traditional estimators such as PostgreSQL across diverse workloads and settings remains challenging.