SRCE: A Workload-Shift-Robust Cardinality Estimation Framework

Cardinality estimation is crucial for query optimization, as it directly affects plan selection and query cost. Traditional methods based on simplified assumptions fail to capture complex correlations in real-world data. Existing learning-based cardinality estimators are typically trained on fixed query distributions and tend to overfit historical predicate combinations and join patterns, leading to severe accuracy degradation when facing unseen workloads.To address this challenge, we propose SRCE, a cardinality estimation framework explicitly designed to be robust to workload shift. SRCE tackles workload shift from both semantic and structural perspectives. It leverages the monotonicity property of query predicates to generate logic-consistent supervision signals, enabling low-cost training data augmentation without additional query executions and improving generalization to unseen predicate combinations. Additionally, SRCE employs graph convolutional networks to model join structures and initializes join node with histogram-based statistics, allowing the estimator to generalize across previously unseen join patterns. To further enhance the model’s robustness, a self-attention mechanism is introduced to perform query-context-aware, dimension-wise reweighting of the concatenated query representation. Extensive experiments on real-world benchmarks demonstrate that SRCE consistently outperforms existing methods, delivering superior estimation accuracy and query performance across both conventional and workload-shifted scenarios.