Automated Machine Learning (AutoML) deals with finding well-performing machine learning models and their corresponding configurations without the need of machine learning experts. However, if one assumes an online learning scenario, where an AutoML instance executes on evolving data streams, the question for the best model and its configuration with respect to occurring changes in the data distribution remains open. Algorithms developed for online learning settings rely on few and homogeneous models and do not consider data mining pipelines or the adaption of their configuration. We, therefore, introduce EvoAu-toML, an evolution-based online learning framework consisting of heterogeneous and connectable models that supports large and diverse configuration spaces and adapts to the online learning scenario. We present experiments with an implementation of EvoAutoML on a diverse set of synthetic and real datasets, and show that our proposed approach outperforms state-of-the-art online algorithms as well as strong ensemble baselines in a traditional test-then-train evaluation.