Web services or other third party or legacy software components which come without code and/or appropriate documentation, are intrinsically tied to the modern an increasingly popular orchestration-based development style of service-oriented solutions. [Active] automata learning has shown to be a powerful means to overcome the perhaps major drawback of these components, their inherent black box character. The success story began a decade ago, when its application led to major improvements in the context of regression testing. Since then, the technology has undergone an impressive development, in particular concerning the aspect of practical application.

The talk will review this development, while focussing on the treatment of data, the major source of undecidability, and therefore the problem with the highest potential for tailored, application-specific solutions.
In the first practical applications of active learning, data were typically simply ignored or radically abstracted. In the meantime, extension to data languages have been developed, which led to the introduction of more expressive models like the so-called register automata. They are able to faithfully represent interface programs, i.e. programs describing the protocol of interaction with components and services. We will illustrate along a number of examples that they can be learned rather efficiently, and that their potential concerning both increased expressivity of the model structure and scalability is high.