Modular Consensus Algorithms for the Crash-Recovery Model

In the crash-recovery failure model of asynchronous distributed systems, processes can temporarily stop to execute steps and later restart their computation from a predefined local state. The crash-recovery model is much more realistic than the crash-stop failure model in which processes merely are allowed to stop executing steps. The additional complexity is reflected in the multitude of assumptions and the technical complexity of algorithms which have been developed for that model. We focus on the problem of consensus in the crash-recovery model, but instead of developing completely new algorithms from scratch, our approach aims at re-using existing crash-stop consensus algorithms in a modular way using the abstraction of failure detectors. As a result, we present three new consensus algorithms for the crash-recovery model for different types of assumptions.