Autonomous Database Management at Scale: Automated Tuning, Performance Diagnosis, and Resource Decentralization

Database administration has always been a challenging task, and is becoming even more difficult with the rise of public and private clouds. Today, many enterprises outsource their database operation to cloud service providers (CSPs) in order to reduce operating costs. CSPs, now tasked with managing an extremely large number of database instances, cannot simply rely on database administrators. In fact, humans have become a bottleneck in the scalability and profitability of cloud offerings. This has created a massive demand for building autonomous databases—systems that operate with little or zero human supervision. While autonomous databases have gained much attention in recent years in both academia and industry, many of the existing techniques remain limited to automating parameter tuning, backup/recovery, and monitoring. Consequently, there is much to be done before realizing a fully autonomous database. This dissertation examines and offers new automation techniques for three specific areas of modern database management. 1. Automated Tuning – We propose a new generation of physical database designers that are robust against uncertainty in future workloads. Given the rising popularity of approximate databases, we also develop an optimal, hybrid sampling strategy that enables efficient join processing on offline samples, a long-standing open problem in approximate query processing. 2. Performance Diagnosis – We design practical tools and algorithms for assisting database administrators in quickly and reliably diagnosing performance problems in their transactional databases. 3. Resource Decentralization – To achieve autonomy among database components in a shared environment, we propose a highly efficient, starvation-free, and fully decentralized distributed lock manager for distributed database clusters.PhDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/153349/1/dyoon_1.pd