Title:Multi-Version Coding in Distributed Storage

Abstract:In applications of distributed storage systems to distributed computing and implementations of key-value stores, the data stored typically changes, and a client desires to read the latest version of the data. We introduce an information theoretic formulation called the multi-version coding problem to study the storage costs incurred in such systems. We consider a storage system with 'n' server nodes. There are 'v' message versions. Each server receives message values corresponding to some arbitrary subset of the v versions, and encodes the messages that it receives. The goal of the multi-version coding problem is to design server encoding functions which enable the following decoding criterion: from any set of c nodes (c \leq n), the value corresponding to the latest common version among this set of c nodes or some version later than that, as per the version ordering, is decodable. We aim to design multi-version codes that minimize the storage cost.
We present two main results in this paper. In our first result, we provide code constructions that achieve smaller storage cost compared to straightforward schemes. In particular, if v|(c-1), the cost is v/(c+v-1), measured in terms of the size of one version. Our code constructions are simple in that we separately encode each version using linear codes. In our second result, we provide an information theoretic lower bound on the storage cost, that is, we show that the storage cost is always lower bounded by an expression that approaches v/(c+v-1) as the version size grows. A consequence of our results is that our code construction is asymptotically optimal as the size of each version grows, when v|(c-1). From a conceptual standpoint, our converse argument for the storage cost lower bound requires interesting manipulations of commonly used converse techniques in network information theory because of an uncommon decoding constraint.