CSR: Small: Efficient Techniques for Modular Past State Systems

回溯是数据存储在数据存储的一致的过去状态上运行特定程序的能力，就好像它们是当前状态一样。回顾使分析过去的状态变得更容易，为审计师、历史学家、经济学家、社会科学家和其他需要调查历史数据的人提供了一个有价值的工具。回顾对于那些想要分析过去状态以预测未来的人来说也很有价值，这是现代数据管理应用程序中越来越需要的功能。关键原因是，现有的回溯技术，为了性能，需要对数据存储内部进行侵入性的难以采用的修改。如果不支持回溯，应用程序开发人员可能很难重构与过去感兴趣的事件相对应的一致状态，本项目将开发一种易于采用的模块化方法和一组相关技术，用于支持轻量级事务数据存储中的回溯，使用嵌入式持久一致过去状态系统。技术挑战是：1）如何在不损害数据存储性能的情况下提供一致的过去状态？过去状态系统需要紧密集成以提高效率，但对内部数据存储组件进行大规模修改是不可行的，需要在数据存储软件堆栈中的低级别运行的新模块化技术。2）如何在跨越大时间间隔的过去状态上高效运行程序？过去的状态需要增量地创建以避免中断数据存储，但是在增量状态上运行程序可能会很慢，需要针对增量数据优化的新的集群和缓存技术。3）可以避免减慢不使用回溯的程序吗？为了评估任何额外的开销，该项目将在工业强度数据存储中开发一个实验原型，并通过实验和分析来回答这个问题。

Retrospection is the ability of a data store to run ad-hoc programs over consistent past states of a data store as if they were the current state. Retrospection makes it easier to analyze past states providing a valuable tool for auditors, historians, economists, social scientists and others with a need to investigate historical data. Retrospection is also valuable to those who want to analyze past states to predict the future, an increasingly in-demand feature in modern data management applications.Most light-weight data stores today do not support retrospection. The key reason is that existing retrospection techniques, for performance, require invasive hard-to-adopt modifications to data store internals. Without support for retrospection, it may be hard for application developers to reconstruct the consistent states corresponding to past events of interest.This project will develop an easy-to-adopt modular method and a set of associated techniques for supporting retrospection in light-weight transactional data stores using anembedded persistent consistent past-state system. The technical challenges are:1) How to provide consistent past states without harming data store performance? The past state system needs to be tightly integrated for efficiency but extensive modifications to the internal data store components are infeasible, requiring new modular techniques that operate at a low-level in the data store software stack.2) How to run programs efficiently over past state that spans large time intervals? past state needs to be created incrementally to avoid disrupting the data store but running a program over incremental state can be slow, requiring new clustering and caching techniques optimized for incremental data.3) Can one avoid slowing down programs that do not use retrospection? To evaluate any additional overhead, the project will develop an experimental prototype in an industrial strength data store, and conduct studies to answer this question experimentally and analytically.