A novel design to support skyline query in key-value stores

Skyline query processing (SQP) is an essential technology for data processing in decision making. Due to the complexity of software development, it is a challenge to implement SQP in key-value stores which have been deployed before. Previous work proposed either to rebuild a new underlying infrastructure or to operate SQP over a centralized management. However, the cost to build a new system from scratch is the obstacle of software development and the single point of failure in a centralized system reduces the system availability. In this paper, we propose two designs of SQP by only using two standard operations in key-value stores, namely, PUT() and GET(), so that our proposal can be built on top of any deployed key-value stores. We use PHTs [1], a distributed index data structure, to exploit the functionalities of range query and k-nearest neighbors query in key-value stores. Our simulations show that RQ-SkyIDX provides excellent performance in a uniform data distribution. On the other hand, KNN-SkyIDX shows lower message overhead than RQ-SkyIDX and exhibits load balance on message overhead in non-uniform data distributions. Through the experiments, we also identify that our proposals can be optimized by a random sampling data set.