Refining the window size of sliding window operations in running data processing systems

In spite of continuous improvements at running production systems, down times have to be reduced further. The challenge is to find means to implement changes online which denotes avoid stopping and restarting the system. As changes on running systems are difficult to handle and may result in lost states a solution is to replicate the running system, modify it in parallel and reintegrate the changes reusing the previous states. In this paper we present an approach of replicating a running data processing system with the special focus on the state retaining reintegration of the modified algorithms. Recovery mechanisms for sliding window operations are proposed and investigated by means of simulation.