Abstract
We present a recursive approach to the design and operation of complex, data-driven IoT services, which are challenged by the traditional cloud-centric view of service delivery. To this end, we introduce a recursive IoT node abstraction, around which we sketch the architecture of a distributed runtime environment for the execution of such services, promoting self-adaptation to a first-class concept. Our framework aims to support services of arbitrary structure and complexity, dynamically configured and fluidly deployed along a heterogeneous device-to-cloud compute continuum, and operating according to provider and user intent. It features inherent privacy enhancements and aims to give users more control over their resources and data. We demonstrate how these principles can be applied to facilitate the composition, deployment and orchestration of complex machine learning workflows, which are becoming increasingly relevant for IoT applications.