Quantifying Eventual Consistency For Aggregate Queries

With the advent of inexpensive cloud computing resources, scalable distributed data stores have surged in popularity [7, 10, 16, 17, 20]. Such systems focus on horizontal scalability and take advantage of cheap, pay by the hour, compute nodes provisioned through the cloud [6]. In doing so, these systems are able to distribute query and insert load across many “shared nothing” compute nodes, improving latency and throughput performance. Consequently, the use of multiple compute nodes increases the likelihood that a node may fail at a given time, making availability a critically important quality [10]. Key-value stores typically address this problem by maintaining � redundant replicas of its data set [10, 16]. In doing so, if a single node in the system fails, � − 1 nodes replicating the same data remain accessible. Increasing � increases the availability of a system. However, introducing redundant replication to a system introduces the problem of consistency. Since networks are unpredictable, each insert operation will arrive at the � different replicas at different times. This leads to the data