Meta-Learned Index Selection via Cross-Workload Transfer Learning

While Wang et al.'s (2025) LITune uses DRL for end-to-end tuning, it requires extensive online training per workload. This idea proposes a meta-learning approach that generalizes across workloads. By training a meta-model on a corpus of workload-index performance pairs (e.g., from Lan et al.'s 2023 disk-resident evaluations), the system can predict optimal index types (ALEX, PGM, etc.) and hyperparameters for unseen workloads in seconds. This extends Choi et al.'s (2024) sampling trade-offs by optimizing not just build-time but also workload-specific selection. Unlike LSIR's model reuse (Patil & Ravishankar 2024), which focuses on spatial data, we target cross-domain transfer (e.g., from numeric to string workloads). The key innovation is using workload embeddings (e.g., data distribution sketches) to query a "knowledge base" of past optimizations, reducing tuning time by 90% compared to LITune.

References:

1. Updatable Learned Indexes Meet Disk-Resident DBMS - From Evaluations to Design Choices. Hai Lan, Z. Bao, J. Culpepper, Renata Borovica-Gajic (2023). Proc. ACM Manag. Data.
2. Can Learned Indexes be Built Efficiently? A Deep Dive into Sampling Trade-offs. Minguk Choi, Seehwan Yoo, Jongmoo Choi (2024). Proc. ACM Manag. Data.
3. Model Reuse in Learned Spatial Indexes. Mayur Patil, Chinya V Ravishankar (2024). International Conference on Statistical and Scientific Database Management.
4. A New Paradigm in Tuning Learned Indexes: A Reinforcement Learning Enhanced Approach. Taiyi Wang, Liang Liang, Guang Yang, Thomas Heinis, Eiko Yoneki (2025). Proc. ACM Manag. Data.