Rethinking Transaction Recovery for High Availability and Performance

点击翻译按钮获取中文摘要

The emergence of many-core architectures and non-volatile memory (NVM) motivates a reconsideration of established techniques for transaction recovery. This recently led to a multitude of designs tailored to modern hardware platforms. Such designs are often incompatible with and more complex than traditional disk-based designs. For instance, recent NVM-centric approaches usually do not support media recovery, which is essential for reliable data services. Other approaches rely on specialized hardware, such as tailor-made I/O controllers or special CPU instructions.In this project, we aim to revisit classical principles of transaction recovery and, by carefully changing fundamental assumptions, devise simple algorithms that work on both traditional and modern architectures.Such new algorithms are able to exploit the performance benefits of modern computer architectures without sacrificing crucial features such as media recovery or requiring specialized hardware.Our approach is based on two families of algorithms: Instant Recovery and REDO-only Recovery. The former is a project initiated at HP Labs and currently being developed in a collaboration with our research group.Instant Recovery aims to drastically reduce the mean time to repair of database systems by supporting incremental, on-demand recovery of data pages while transactions are running.In this project proposal, we aim to continue this work by implementing Instant Restore, which applies Instant Recovery ideas to media failures.The latter family of algorithms has been conceptualized by our research group and we aim to implement and evaluate them in the context of this project. The goal is to devise new strategies for commit processing and propagation control, leading to "REDO-only" recovery algorithms which can be easily applied to distributed logging and non-volatile memories, thus exploiting the full potential of modern hardware architectures.Our final goal is to combine these two families of algorithms into a system design for efficient, robust, highly available, and hardware-transparent transaction processing.