Adaptive query execution for data management in the cloud

A major component of many cloud services is query processing on data stored in the underlying cloud cluster. The traditional techniques for query processing on a cluster are those offered by parallel DBMS. These techniques however, cannot guarantee high performance for cloud; parallel DBMS lack adequate fault tolerance mechanisms in order to deal with non-negligible software and hardware failures. MapReduce, on the other hand, allows query processing solutions that are fault tolerant, but imposes substantial overheads. In this paper, we propose an adaptive software architecture which can effortlessly switch between MapReduce and parallel DBMS in order to efficiently process queries regardless of their response times. Switching between the two architectures is performed in a transparent manner based on an intuitive cost model, which computes the expected execution time in presence of failures. The experimental results show that the adaptive architecture achieves the lowest possible query execution time for various scenarios.