Techniques are disclosed for validating the resiliency of a networked application made available using a collection of interacting servers. In one embodiment, a network monitoring application observes each running server (or application) and at unspecified intervals, picks one and terminates it. In the case of a cloud based deployment, this may include terminating a virtual machine instance or terminating a process running on the server. By observing the effects of the failed server on the rest of the network application, a provider can ensure that each component can tolerate any single instance disappearing without warning.