A Consistent Replica Selection Approach for Distributed Key-Value Storage System

When the amount of data is increasing, most of the distributed key-value storage systems use the quorum based consistent replica selection algorithms for fault-tolerance and load balancing. These algorithms allow higher consistency because the quorum requires a majority of replicas to execute the request. The fewer failure nodes the system has, the higher the consistency. On the other hand, it has a high latency of a read/write request due to the static consistency level for a quorum. Higher read/write consistency level dramatically increases the read/write latency and is not suitable for accessing the updated version of the data and massive workload. We propose a consistent replica selection approach with dynamically changing the consistency level. The system distributes data in a distributed hash table for the write request. And the system searches the nearest replica with an existing consistent hashing algorithm instead of random selection and predicts the staleness rate with the Probability of Bounded Staleness (PBS) to select the consistent replicas for the read request. The experimental result shows that the proposed system reduces the staleness rate and the number of replicas by varying consistency level.