The current architecture works well but connection with the scheduler may get lost after a long period of network inactivity.

The task here is to refactor the architecture so that adding and removing node as well as crashing nodes gets improved handling and thus avoids as much as possible "blocked" experiments.