Query Optimisation

Cardinality Estimation for Similarity Search on High-Dimensional Data Objects: The Impact of Reference Objects (VLDB'25) (Code)

Hai Lan, Shixun Huang, Zhifeng Bao, Renata Borovica-Gajic

This paper studies cardinality estimation for similarity search on high-dimensional data (CE4HD), which is essential for query optimization in vector databases and embedding-based applications. It proposes a novel reference-object–based approach, where selected data objects with similar cardinality patterns are used to estimate the query cardinality efficiently and accurately. The paper introduces two methods—SRCE (single-reference) and MRCE (multi-reference)—that leverage reference objects and lightweight models to improve robustness and efficiency. Experiments show that the proposed methods achieve up to 10× faster estimation and up to 136× lower Q-error compared to prior approaches such as SimCard and SelNet.

Practical Parameterized Query Optimization via Efficient Plan Reuse and List-wise Ranking (SIGMOD'26) (Code)

Hai Lan, Yang Yu, Zhifeng Bao, Zi Huang, Yuwei Peng

PLARQ is a practical learned optimizer designed for parameterized query optimization (PQO), where the same query template is repeatedly executed with different parameter values. It introduces a plan reuse strategy that retrieves high-quality candidate plans from a precomputed pool based on query similarity, and a list-wise attention-based ranking model to select the best plan efficiently. PLARQ integrates seamlessly with PostgreSQL without modifying optimizer internals and achieves significant performance improvements, with up to 420× speedup over PostgreSQL and up to 2× over existing learned methods.

A Survey on Advancing the DBMS Query Optimizer: Cardinality Estimation, Cost Model, and Plan Enumeration (Data Science and Engineering'21)

Hai Lan, Zhifeng Bao, Yuwei Peng

This paper surveys techniques for improving cost-based database query optimizers, focusing on three core components: cardinality estimation, cost models, and plan enumeration. It explains that inaccuracies in cardinality estimation are the main cause of suboptimal query plans, and reviews traditional methods (e.g., histograms and sampling) as well as modern learning-based approaches. The paper also highlights key limitations of existing optimizers and outlines future directions, especially integrating machine learning to build more accurate and robust optimization frameworks.

An Index Advisor Using Deep Reinforcement Learning (CIKM'20) (Code)

Hai Lan, Zhifeng Bao, Yuwei Peng

This paper studies the index selection problem, which aims to choose the best set of indexes to minimize workload execution cost under storage or number constraints, a known NP-hard problem. It proposes a deep reinforcement learning–based index advisor that integrates heuristic candidate generation with a Deep Q Network (DQN) to recommend both single-attribute and multi-attribute indexes while modeling their interactions. The approach enables more effective exploration of index combinations and supports multiple-index access to tables. Experiments show that the proposed method achieves better workload performance than traditional greedy and optimization-based index advisors.