In recent years, publish/subscribe (pub/sub) systems have evolved as serious candidates for the implementation of multicast communication. However, to date they have not been properly analysed with respect to the critical dimension of robustness. Indeed, the robustness evaluations found in the literature only cover a subset of the types of failure situations and generally lack a comprehensive architectural classification. In this context, this paper introduces the necessary foundations for systematic evaluation of the robustness of pub/sub system: a comprehensive classification scheme and a complete taxonomy of failures. We argue that both robustness evaluation of pub/sub systems and development methods of robust pub/sub systems should take rational behavior into account. We demonstrate this point through an experimental study.