Survey on Learnable Databases: A Machine Learning Perspective

During the decades of development of artificial intelligence, a spectrum of applications involving image, speech, text data, etc. are successfully powered by machine learning. The advantages are mainly derived from the learning ability from data, which most traditional databases lack. Recent years have seen a surge in approaches that explore artificial intelligence to power traditional databases, i.e. learnable databases, making databases more adaptive and intelligent. Specifically, they can be automatically optimized according to historical metric statistics and current query workload, which significantly improves the database performance and relieves the trivial routine maintenance suffering. However, one of the major issues, especially for practitioners, is the lack of consensus in their definitions as well as a lack of clear categorization from a machine learning perspective. To alleviate these problems, this paper introduces concepts and algorithms related to learnable databases and investigates the progress in learnable databases in five aspects: database parameter configuration, data storage management, query optimization, query interface, and benchmark of learnable databases. Additionally, we survey AI-empowered technique development in commercial databases and new approaches to learning-based database security. We develop a categorizing framework in terms of input features, model selection, and output results (mostly being viewed as class labels). Finally, we conclude the current work and discuss future work on learnable databases.