CaDRoP: Cost Optimized Convergent Causal Consistency in Social Network Systems
Ta-Yuan Hsu, Ajay D. Kshemkalyani · 2021
Asynchronous geo-replication for data resources is used to provide high availability and lower latency in modern cloud store systems. Convergent causal consistency is the corner-stone to provide useful semantics for online human interaction services. Compared to full replication, partial replication has potential benefit of lower message counts in social network systems. However, static replication is ineffective for time-varying workloads. We propose a causal+ consistency protocol, CaDRoP, to support dynamic replication, and ensure the convergence property for all comments following a post and the causal ordering between posts with explicit causality. We evaluate CaDRoP protocol with realistic workloads by different PUT rates in terms of the practical price of Amazon AWS. The results show that CaDRoP incurs much lower cost than the statically replicated data store in another causal+ algorithm. We further evaluate CaDRoP by comparing it with a clairvoyant optimal replication solution. The findings indicate that with cache, CaDRoP incurs only around 6% ~ 16% extra cost. Without cache, CaDRoP brings around 2% ~ 4.5% extra cost in steady states.