A New Paradigm in Tuning Learned Indexes: A Reinforcement Learning Enhanced Approach

TL;DR

This work addresses the challenging problem of tuning learned index structures (LIS) to workloads and distributions. It introduces LITune, an end-to-end automatic tuner that combines a tailor-made Deep Reinforcement Learning pipeline with an online-offline (O2) updating mechanism and a safety-focused ET-MDP framework. Through offline meta-trained initialization and online adaptation, LITune achieves substantial runtime reductions and throughput gains, while maintaining safety during exploration and rapid responsiveness to data distribution shifts. The approach is validated on multiple LIS instances and workloads, demonstrating strong adaptability, efficiency, and robustness with significant improvements over traditional tuning methods. The findings suggest LITune can significantly broaden the practical deployment of LIS in real-world systems by automating complex tuning under dynamic conditions.

Abstract

Learned Index Structures (LIS) have significantly advanced data management by leveraging machine learning models to optimize data indexing. However, designing these structures often involves critical trade-offs, making it challenging for both designers and end-users to find an optimal balance tailored to specific workloads and scenarios. While some indexes offer adjustable parameters that demand intensive manual tuning, others rely on fixed configurations based on heuristic auto-tuners or expert knowledge, which may not consistently deliver optimal performance. This paper introduces LITune, a novel framework for end-to-end automatic tuning of Learned Index Structures. LITune employs an adaptive training pipeline equipped with a tailor-made Deep Reinforcement Learning (DRL) approach to ensure stable and efficient tuning. To accommodate long-term dynamics arising from online tuning, we further enhance LITune with an on-the-fly updating mechanism termed the O2 system. These innovations allow LITune to effectively capture state transitions in online tuning scenarios and dynamically adjust to changing data distributions and workloads, marking a significant improvement over other tuning methods. Our experimental results demonstrate that LITune achieves up to a 98% reduction in runtime and a 17-fold increase in throughput compared to default parameter settings given a selected Learned Index instance. These findings highlight LITune's effectiveness and its potential to facilitate broader adoption of LIS in real-world applications.

Figures (12)

• Figure 1: (a) shows the performance surface of a learned index (ALEX) under a wild exploration of the parameter space. (b) highlights the optimal performance speedup achieved by LITune compared to default expert-selected parameters. (c) illustrates the continuous tuning performance of our system alongside other out-of-the-box methods under default configurations. (d) compares the tuning stability and costs across methods to reach their respective optimal performance levels.
• Figure 2: Selected parameter value distributions and their impact scores across different workloads when tuning on ALEX. The heatmap colors represent normalized optimal parameter values, while the percentages indicate each parameter's individual tuning impact.
• Figure 3: The architecture of LITune. Part A illustrates the training phase, where RL-based models are trained. Once the training is complete, these models are deployed as online tuners in Part B. The operational details of the O2 system are explained in Part C.
• Figure 4: (a) Running example of LITune. This example demonstrates how LITune's tuner components respond to changes in performance metrics and adjust to workload shifts. (b) The safe-RL approach prevents aggressive tuning by learning from instabilities encountered during training.
• Figure 5: Tuning efficiency--Performance as tuning steps increase. Above: runtime ratio (best found vs. default settings). Below: throughput ratio (best found vs. default settings).

Theorems & Definitions (4)

• definition 1: Constrained Markov Decision Process
• Remark
• definition 2: Early Terminated MDP
• Remark