How to effectively and efficiently monitor, manage, and adapt web services in a composite service or a service-oriented application is becoming a significant issue. In this paper, we argue that it is insufficient to only solve emerging service faults at the deployment time or runtime; instead, we propose that the prediction of service faults is equally important. We devised a risk-driven service composition management (RDSCM) approach including four main phases: (1) preparation, (2) planning, (3) monitoring and reaction, and (4) analysis. By applying the proposed approach, risky component services can be removed earlier, and the fault source can be tracked and identified more easily when a fault occurs. We developed a prototype to realize the proposed approach, and conducted experiments to verify the approach. The implementation and experiments demonstrate that the proposed risk-driven approach can effectively and efficiently ensure the robustness of a service-oriented system.