A Modular Approach for Non-Distributed Crash Recovery for Streaming Systems
ISSN der Zeitschrift
Datenbanksysteme für Business, Technologie und Web (BTW 2017)
Streaming and Dataflows
Gesellschaft für Informatik, Bonn
High availability and reliability are important aspects for streaming systems. State-of-the-art recovery techniques like active or passive standby use several nodes to fulfill these requirements. But not every processing task is done in a professional environment with resources for a distributed system (e.g. smart home). Additionally, even distributed streaming systems can benefit from reliable non-distributed crash recovery (NDCR) because it can help to restore the overall system state faster not only after a node failure but also after the roll-out of updates. In this paper, we present our research on NDCR for streaming systems and point out its possibilities and limitations. We propose a flexible and extensible framework in which small NDCR tasks can be combined to high-level NDCR classes with different degrees of fulfillment for completeness and correctness: at-most-once, at-least-once or exactly-once. Further, we offer a way to mark elements that may be incorrect or duplicated.