We consider scheduling real-time distributable threads in the presence of node/link failures, message losses, and dynamic node joins and departures. We present a distributed scheduling algorithm called RTMG. The algorithm uses gossip-based communication for discovering eligible nodes. Traditionally, gossip protocols incur high message overhead. We explain that this problem is not that serious. We present a hybrid message propagation protocol with lower message overhead, and improve it by evenly distributing the overhead into all gossip rounds. In scheduling local thread sections, RTMG exploits slacks to optimize gossip time utilization. Thereby, it satisfies end-to-end time constraints with probabilistic assurance. Our simulation studies verify our analytical results.