Benchmarking Learned Cardinality Estimation Techniques for Analytical Query Processing in Data Warehouses

Cardinality estimation remains one of the most critical yet error-prone components of query optimization in modern data warehouses. Recent advances in machine learning have produced a diverse family of learned cardinality estimators that demonstrate substantial accuracy improvements on standard benchmarks. Yet existing evaluations predominantly rely on third-normal-form schemas, leaving their effectiveness on star and snowflake schemas—the backbone of analytical data warehousing—largely unexplored. This paper presents a systematic empirical evaluation of seven representative learned cardinality estimation methods spanning three paradigmatic categories: query-driven, data-driven, and hybrid approaches. All methods are benchmarked against the PostgreSQL histogram-based estimator on three complementary datasets: TPC-DS with its native snowflake schema, STATS-CEB with real-world relational data, and IMDB/JOB as the established cross-study reference. The evaluation encompasses estimation accuracy measured by Q-Error and P-Error, inference latency, training cost, model compactness, end-to-end query execution time, and robustness under simulated ETL batch insertions. Results indicate that hybrid methods, particularly FactorJoin, achieve the strongest accuracy on data warehouse workloads with a median Q-Error of 1.74 on TPC-DS, while data-driven methods such as FLAT and BayesCard offer a favorable balance between accuracy and inference speed. BayesCard and FactorJoin exhibit the highest resilience to data updates, with median Q-Error increasing by fewer than 1.5 points after a 50% data insertion. These findings provide actionable guidance for practitioners seeking to deploy learned cardinality estimation in production data warehouse environments.