Towards Resilient Execution of Adaptation in Decentralized Self-Adaptive Software Systems

In an environment of ever-growing numbers of IoT devices and distributed software systems, we face particular challenges regarding updates and maintenance of the underlying system. Generally, this maintenance of all nodes would require human interaction and is costly and time-consuming. A solution to that are self-adaptive software systems (SAS) [1], [2], which handle the growing complexity of distributed software systems by enabling run-time adaptation. Decentralized solutions [3] for SAS (dSAS) have been developed recently to eliminate single points of failures and bottlenecks compared to centralized approaches where the adaptation logic relies on a single node. However, dSAS require a sophisticated coordination mechanism to ensure a reliable adaptation process on multiple, decentralized devices. Existing approaches for the coordination of the adaptation execution [4]–[6] are based on concepts of the Two-Phase-Commit (2PC) protocol [7] to execute structural adaptations which modify the systems architecture depending on the current context. They ensure a consistent outcome of the adaptation on all nodes - either all adaptations are performed or rolled back - and thus, in any case, a consistent system configuration. In case of an error or disconnection during the adaptation, intermediate changes and modifications get reverted to their original state. This imposes a drawback for self-optimizing (i.e., increasing performance, efficiency, and quality) [8] adaptations, where the system should work even if one node fails to adapt. A rollback is non-essential when the whole system can continue working even if the execution failed for parts of the system. In our system model, this partial adaptation improves the overall task execution. Our work aims to tackle those issues of a strict transactional approach for adaptation by introducing a relaxed consistency constraint for dSAS. We present a protocol for coordinating adaptation operations performed by a set of distributed nodes that allows the different outcomes of participating nodes. Nodes detect inconsistencies and aim to recover when the failed node is ready again.