Separating durability and availability in self-managed storage

Building reliable data storage from unreliable components presents many challenges and is of particular interest for peer-to-peer storage systems. Recent work has examined the trade-offs associated with ensuring data availability in such systems. Reliability, however, is more than just availability. In fact, the durability of data is typically of more paramount concern. While users are likely to tolerate occasional disconnection from their data (they will likely have no choice in the matter), they demand a much stronger guarantee that their data is never permanently lost due to failure. To deliver strong durability guarantees efficiently, however, requires decoupling durability from availability. This paper describes the design of a data redundancy scheme that guarantees durability independently from availability. We provide a formula for determining the rate of redundancy repair when durability is the only concern and show that availability requires much more frequent repair. We simulate modified versions of the Total Recall block store that incorporate our design. Our results show that we can deliver durability more cheaply than availability, reducing network overhead by between 50% and 97%.