TimeLine: A High Performance Archive for a Distributed Object Store

This paper describes TimeLine, an efficient archive service for a distributed storage system. TimeLine allows users to take snapshots on demand. The archive is stored online so that it is easily accessible to users. It enables "time travel" in which a user runs a computation on an earlier system state.

Archiving is challenging when storage is distributed. In particular, a key issue is how to provide consistent snapshots, yet avoid stopping user access to stored state while a snapshot is being taken. The paper defines the properties that an archive service ought to provide and describes an implementation approach that provides the desired properties yet is also efficient. TimeLine is designed to provide snapshots for a distributed persistent object store. However the properties and the implementation approach apply to file systems and databases as well.

TimeLine has been implemented and we present the results of experiments that evaluate its performance. The experiments show that computations in the past run well when the archive store is nearby, e.g., on the same LAN, or connected by a high speed link. The results also show that taking snapshots has negligible impact on the cost of concurrently running computations, regardless of where the archived data is stored.