CSR: Medium: Eidetic Systems

一台普通计算机产生的绝大多数状态都是生成、消耗、然后永远丢失的。丢失状态包括进程地址空间、删除的文件、进程间通信和从网络接收的输入。状态丢失意味着价值丢失：用户无法恢复有关过去计算的详细信息，这些信息对于审计、取证、调试、错误跟踪和许多其他目的是有用的。为了解决这个问题，这个项目正在开发能够根据需要召回计算机上存在的任何过去状态的特征计算机系统，包括所有文件的所有版本、瞬态应用程序状态和网络通信。此外，智能计算机系统可以以字节粒度解释当前和过去状态的来源：例如，它们可以回答诸如“此数据来自何处以及使用哪些步骤来派生数据？”或“此数据影响了什么状态或输出？”建议的工作是部署几个重要的系统特性，利用特征系统：(1)完全召回：计算机系统不仅可以恢复保存在磁盘上的任何先前数据（如版本控制文件系统），还可以恢复任何先前的应用程序状态、瞬态输出或计算状态。(2)完整的来源：对于任何数据对象，系统可以提供该数据如何产生的历史，包括进程间通道和转换数据的计算。(3)丰富的查询：用户可以对整个执行历史进行推理，以检测异常，恢复工作流，提高生产力。(4)重放结构化存储：可以通过在端点确定性地重新生成文件数据来减少网络使用，并且可以通过机会性地重复删除非确定性输入而不是文件数据的日志来减少存储使用。(5)保护隐私的重放：来源可以实现全面的删除策略，其中所有来自敏感数据的值都被识别和编辑。

The vast majority of state produced by a typical computer is generated, consumed, and then lost forever. Lost state includes process address spaces, deleted files, interprocess communication, and input received from the network. With lost state comes lost value: users cannot recover detailed information about past computations that would be useful for auditing, forensics, debugging, error tracking, and many other purposes.To solve this problem, this project is developing eidetic computer systems that can recall, on demand, any past state that existed on a computer, including all versions of all files, transient application state, and network communication. Further, eidetic computer systems can explain the provenance of current and past state at byte granularity: for example, they can answer questions such as "Where did this data come from and what steps were used to derive the data?" or "What state or outputs did this data influence?"The proposed work is deploying several important system features that utilize eidetic systems: (1) Total recall: a computer system can recover not just any prior data saved to disk (as in versioning file systems), but also any prior application state, transient output, or computation state. (2) Complete provenance: for any data object, the system can provide the history of how that data was produced, including inter-process channels and the computations that transformed the data. (3) Rich queries: users can reason over the entire history of execution to detect anomalies, recover workflows, and improve productivity. (4) Replay-structured storage: network usage can be reduced by deterministically regenerating file data at endpoints, and storage usage can be reduced by opportunistically deduplicating logs of non-deterministic inputs rather than file data. (5) Privacy-preserving replay: provenance can enable comprehensive deletion policies in which all values derived from sensitive data are identified and redacted.