Minimizing coordination in replicated systems

Abstract

Replication has been widely adopted to build highly scalable services, but this goal is often compromised by the coordination required to ensure application-specific properties such as state convergence and invariant preservation. In this paper, we propose a principled mechanism to minimize coordination in replicated systems via the following components: a) a notion of restriction over pairs of operations, which captures the fact that the two operations must be ordered w.r.t. each other in any partial order; b) a generic consistency model which, given a set of restrictions, requires those restrictions to be met in all admissible partial orders; c) principles for identifying a minimal set of restrictions to ensure the above properties; and d) a coordination service that dynamically maps restrictions to the most efficient coordination protocols. Our preliminary experience with example applications shows that we are able to determine a minimal coordination strategy.