Toward Eidetic Distributed File Systems

We propose a new point in the design space of versioning and provenance-aware file systems in which the entire operating system, not just the file system, supports such functionality. We leverage deterministic record-and-replay to substitute computation for data. This leads to a new file system design where the log of nondeterministic inputs, not file data, is the fundamental unit of persistent storage. We outline a distributed storage system design based on these principles and describe the challenges we foresee for achieving our vision.