An Architecture for Dynamic Scalable Self-Managed Persistent Objects

This paper presents a middleware architecture and a generic orchestrating protocol for implementing persistent object operations for large scale dynamic systems in a self-managing manner. In particular, the proposed solution is fully distributed, allows dynamic changes in the environment, and nodes are neither assumed to be aware of the size of the system nor of its entire composition.The architecture includes two modules and three services. The modules are expected to be instantiated and executed among relatively small sets of nodes in the context of a single multi-object operation (operation that spans multiple persistent objects) and therefore, can be implemented using known classical distributed computing approaches. On the other hand, services are long lived abstractions that may involve all nodes and should be implemented using known peer-to-peer techniques. The main contribution of the paper is to provide, for the first time, an architecture that brings together several seemingly distinct research areas, namely distributed consensus, group membership, notification services (publish/subscribe), scalable conflict detection (or locking), and scalable persistent storage. All these components are orchestrated together in order to obtain (strong) consistency on an a priori unsafe system.This paper also promotes the use of oracles as a design principle in implementing the respective components of the architecture. Specifically, each of the modules and services are further decomposed into a “benign” part and an “oracle” part, which are specified in a functional manner. This makes the principles of our proposed solution independent of specific implementations and environment assumptions (e.g., it does not depend on any specific distributed hash tables or specific network timing assumptions, etc). The contribution of this paper is therefore largely conceptual, as it focuses on defining the right architectural abstractions and on their orchestration, rather than on the actual mechanisms that implement each of its components.