Data Organisation

A Fully On-disk Updatable Learned Index (ICDE'24) (Code)

Hai Lan, Zhifeng Bao, J. Shane Culpepper, Renata Borovica-Gajic, Yu Dong

AULID proposes a fully on-disk updatable learned index designed specifically for disk-based database systems, where disk I/O dominates performance. It introduces a disk-efficient structure that reduces I/O cost by minimizing model traversal and optimizing node layouts for disk access. AULID supports efficient lookups, inserts, and updates while maintaining high accuracy and low storage overhead. Experiments show that AULID significantly improves lookup and update performance compared to traditional B+-trees and prior learned indexes in disk-resident settings.

Updatable Learned Indexes Meet Disk-Resident DBMS — From Evaluations to Design Choices (SIGMOD'23) (Code)

Hai Lan, Zhifeng Bao, J. Shane Culpepper, Renata Borovica-Gajic

This paper studies whether updatable learned indexes are effective in disk-resident DBMS environments, where most real-world databases operate. It implements and evaluates several state-of-the-art learned indexes (e.g., PGM, LIPP, ALEX) and compares them against traditional B+-trees under various workloads. The results show that while learned indexes can outperform B+-trees in specific workloads, B+-trees remain highly competitive overall. Based on extensive analysis, the paper identifies key limitations of learned indexes on disk and proposes design principles to guide the development of efficient on-disk learned indexes.

Missing Value Imputation in Tabular Data Lakes Unleashed: A Hybrid Approach (VLDBJ'26) (System Prototype)

Feng Luo, Hai Lan, Hui Luo, Zhifeng Bao, J. Shane Culpepper, Shazia Sadiq, Xiaoli Wang

Missing values in tabular data lakes can severely impact data analysis and diminish the performance in downstream applications. Ideally, a robust imputation strategy should properly take three aspects of variety into consideration: source of imputed value, the types of tables involved, and the data types of the missing value. However, the existing estimation-based or search-based imputation approaches do not effectively account for all three aspects of variety. To address this gap, we propose CESID, a hybrid method that effectively and efficiently combines the strengths of both search-based and estimation-based approaches, enabling its robustness in dealing with the three varieties. Extensive experiments conducted on three data lakes demonstrate that CESID effectively and efficiently addresses the missing value problem.