Preface

Guoliang Li, Dong Deng, Ju FanJCST 2021 · DOI

Abstract

This preface introduces a special section of the Journal of Computer Science and Technology (July 2021) dedicated to the intersection of Artificial Intelligence (AI) and Database (DB) systems. It highlights the bidirectional synergy between these fields, categorized as AI4DB (using machine learning to optimize database internals) and DB4AI (applying database principles to enhance AI efficiency and robustness). The section presents seven peer-reviewed papers that explore diverse methodologies, including deep reinforcement learning, convolutional neural networks, and transfer learning, to address challenges in workload-aware tuning, cardinality estimation, and model management.

Introduction

Historically, database management and artificial intelligence were viewed as separate disciplines. However, the explosion of data volume and the increasing complexity of modern systems have necessitated a convergence. Manual tuning and heuristic-based designs in databases are reaching their limits, while AI models face significant hurdles regarding deployment efficiency and data quality.

Guo-Liang Li, Guest Editor

To bridge this gap, guest editors Guo-Liang Li, Dong Deng, and Ju Fan organized this special section to explore the "virtuous cycle" between AI and databases. This synergy is twofold: AI4DB focuses on automating database components with adaptive models, while DB4AI leverages database expertise to manage the AI lifecycle and handle real-world data imperfections.

Dong Deng, Guest Editor

The Dual Paradigm: AI4DB and DB4AI

The research presented in this section addresses two core pillars:

  1. AI4DB (AI for Databases): This involves replacing traditional "rule-of-thumb" components—such as index structures and query optimizers—with learned models. These models are designed to adapt dynamically to specific workloads, improving performance beyond what static heuristics can achieve.
  2. DB4AI (Databases for AI): This direction applies database techniques to accelerate AI deployment. It focuses on managing model lifecycles, ensuring reusability, and addressing the "messiness" of real-world data through systematic cleaning and quality control.

Methodological Innovations

The seven papers selected from 20 submissions utilize a variety of advanced techniques:

  • Deep Reinforcement Learning (DRL): Applied in systems like WATuning for automated, workload-aware database configuration.
  • Convolutional Neural Networks (CNNs): Used to improve the accuracy of cardinality estimation in SQL query processing.
  • Transfer Learning: Employed to model the performance of distributed graph computations, particularly when training data is limited.
  • Cache-Conscious Design: Integrated into learned index structures (e.g., COLIN) to ensure high performance on modern hardware.
  • Attention Mechanisms: Utilized for both database tuning and deep entity matching to identify records referring to the same real-world objects.

Key Insights and Research Highlights

The special section emphasizes that the symbiosis between AI and databases is a fundamental shift rather than a simple integration. Key takeaways include:

  • Workload Awareness: Static configurations are no longer sufficient. Attention-based DRL allows databases to "understand" and self-optimize based on current workloads.
  • Practicality of Learned Components: Learned indexes must be hardware-aware (cache-conscious) to be viable in production environments.
  • Data Quality as a First-Class Citizen: Empirical studies on "dirty data" demonstrate that data cleaning is a critical bottleneck for ML model performance and must be integrated into the AI pipeline.

Results and Future Directions

The featured research demonstrates significant improvements in performance and robustness. For instance, the COLIN index structure achieves high read/write efficiency, while CNN-based estimators provide more precise results for complex SQL queries. Furthermore, the introduction of Online Machine Learning (OLML) database systems shows that models can be stored and reused efficiently for new tasks.

While this preface focuses on the contributions of the accepted papers, it underscores that the integration of AI and DB is an ongoing, evolving field. The challenge remains to build systems that are not only high-performing but also robust against the complexities of real-world data and dynamic workloads.

Summary updated

BibTeX
@article{li2021preface,
  author = {Guoliang Li and Dong Deng and Ju Fan},
  title = {{Preface}},
  journal = {J. Comput. Sci. Technol.},
  year = {2021},
  doi = {10.1007/s11390-021-0005-0},
}
Figure