Epidemia: Variable consistency for transactional cloud databases

Classic replication protocols running on traditional cluster-based databases are currently unable to meet the ever-growing scalability demands of many modern software applications. Recent cloud-based storage repositories overcome such limita-tions by fostering availability and scalability over data consistency and transactional support. However, many applications that cannot resign from their transactional na-ture are unable to benefit from the cloud paradigm. This paper presents Epidemia, a distributed storage architecture featuring a hybrid approach that combines classic database replication with a cloud-inspired infrastructure to provide transactional sup-port and high availability. This architecture is able to offer different consistency levels according to the client demands, thanks to a replication strategy based on epidemic updates in which the replicas of each data partition are organized hierarchically. Addi-tionally, the behavior of a prototype implementation under different workload scenarios is evaluated. Conducted experiments verify that (1) configuration parameters such as the partitioning scheme or the replication protocol play a crucial role on system's throughput, and (2) the existence of replica hierarchies that are asynchronously up-dated is able to alleviate the scalability limitations of traditional replicated databases by directing transactions that tolerate a certain staleness in the versions of retrieved data items to these replicas.