Resiliency-Aware Data Management

Computing architectures change towards massively parallel environments with increasing numbers of heterogeneous components. The large scale in combination with decreasing feature sizes leads to dramatically increasing error rates. The heterogeneity further leads to new error types. Techniques for ensuring resiliency in terms of robustness regarding these errors are typically applied at hardware abstraction and operating system levels. However, as errors become the normal case, we observe increasing costs in terms of computation overhead for ensuring robustness. In this paper, we argue that ensuring resiliency on the data management level can reduce the required overhead by exploiting context knowledge of query processing and data storage. Apart from reacting on already detected errors, this was mostly neglected in database research so far. We therefore give a broad overview of the background of resilient computing and existing techniques from the database perspective. Based on the lack of existing techniques on data management level, we raise three fundamental challenges of resiliency-aware data management and present example use cases. Finally, our vision of resiliency-aware data management opens many directions of future work. Fundamental research, including the partial reuse of underlying mechanisms, would allow data management systems to cope with future hardware characteristics by effectively and efficiently ensuring resiliency.