Recoverability of Erasure Codes Based and Intensively Written Storage Systems

We consider a mutable distributed storage system that protects data using systematic linear k-of-n codes. Such a system may get into a partial updated state if some failures occur, and may become unrecoverable even though there are k+ living bricks (for they may be inconsistent with each other). General recovery mechanisms judge whether the system is recoverable according to the maximum number of consistent living bricks. We prove the condition that there are at least k consistent living bricks is not necessary for system recovery. We also investigate the recoverability that can be gained under the general recovery mechanisms. The condition for assuring recoverability is figured out, and the quantitative evaluation of stochastic recoverability is explored. Our simulation results show that the tolerable bounds of erased blocks not only depend on the number of partial updates, but also on the proportion between original blocks and redundant blocks that are erased.