Cloud computing has enabled many applications to dynamically accommodate their resources in response to variations in their workloads. Elastic scaling is implemented mostly via reactive techniques that are slow to respond and may induce service degradation during the adaptation period. To avoid those pitfalls, proactive techniques have emerged as an alternative. However, these are typically limited to settings with homogeneous resources. We introduce Augure, a prediction-based controller for live reconfiguration of cloud applications in heterogeneous settings. Augure relies on behavioral patterns of the workload obtained from historical data to feed a proactive adaptation engine. With the help of constraint solvers, Augure's engine finds the combination of (potentially mixed) resource types that best matches the expected evolution of the workload, and derives a plan that minimizes the price billed by the cloud provider and the impact of the reconfiguration on the quality of service provided to clients. We use simulations and a real system implementation to evaluate Augure and compare it to other controllers such as Reactive, Vadara, and Vadara+.