Cluster metadata recovery

摘要

A method and apparatus are described for recovering a fully consistent file system stored in a cluster file system with multiple metadata servers using an epoch of undo records. The epoch consists of (i) a virtual instantaneous snapshot marking a consistent and valid file system image and (ii) a set of undo records that enable the file servers to roll-back to this fully consistent image associated with the file system. The file system is recovered by rolling back file transactions associated with undo records subsequent to the undo records associated with the epoch snapshot. In addition, the undo records are maintained by advancing the epoch value and purging unneeded undo records.