In the context of large decentralized many-to-many communication systems it is impractical to provide realistic and hard bounds for certain QoS metrics including latency bounds. Nevertheless, many applications can yield better performance if such bounds hold with a given probability. In this paper we show how probabilistic latency bounds can be applied in the context of publish/subscribe. We present an algorithm for maintaining individual probabilistic latency bounds in a highly dynamic environment for a large number of subscribers. The algorithm consists of an adaptive dissemination algorithm as well as a cluster partitioning scheme. Together they ensure i) adaptation to the individual latency requirements of subscribers under dynamically changing system properties, and ii) scalability by determining appropriate clusters according to available publishers in the system.