The emergence of Grid architectures brings in new challenges in terms of high latencies, dynamic resource availability, and data sharing. Data Grids connect collections of thousands of geographically distributed computational and storage resources, and promote their sharing between users. Data sharing can be accomplished through data replication. Although this later is a well-established technique, its implementation presents additional problems such as the management of replica consistency. In this paper, we propose a new model to address this specific problem in large scale systems. The model combines two existing approaches. It benefits from their advantages and offers more. It optimizes the cost of replica consistency management and resists to changes in large systems.