This paper presents Pirrus, a replica management system that addresses the problem of providing data availability on a wireless sensor network. Pirrus uses probabilistic failure models (e.g., derived from environmental conditions and estimation of available energy) to adaptively create and maintain a number of replicas of the data. Replica management is formulated as an energy optimization problem, then solved with a greedy heuristic that only uses information gathered from neighbors. Intuitively, Pirrus trades off the energy saved by limiting the number of replicas when the network health is good to extend the lifetime when more replicas are needed. Our simulation results show how the solution provided by Pirrus achieves good performance with a sustainable computational cost. Compared to the performance of a fixed number of replicas, Pirrus extends the network lifetime by more than 20%.