Reliable and automated management technologies are essential to support the fast growth of Internet of Things (IoT) applications and infrastructures. Manually deploying IoT applications on thousands of devices in a heterogeneous environment is complex, time-consuming, and error-prone. IoT devices are mostly embedded systems which are deployed as edge devices at specific physical locations where they provide their service by interacting with the physical environment and each other. For example, outdoor temperature sensors, traffic sensors on highways, or remote controlled lights. From a technical perspective, this cyber-physical nature of IoT applications is their most valuable but also their most challenging characteristic. To keep up with the proliferation of IoT technologies, as well as the fast growing needs of IoT applications, their development and deployment speed must increase accordingly. Techniques such as DevOps and continuous delivery, which are well-known in the context of cloud applications, are slowly adapted for IoT applications. One challenge of this process is to automate the deployment and management of IoT applications on edge infrastructures. The Topology and Orchestration Specification for Cloud Applications (TOSCA) enables the automated provisioning and management of various kinds of applications. However, its general-purpose modeling language makes it difficult to capture the cyber-physical nature of IoT applications. Existing TOSCA orchestration engines do not account for the low reliability, size, and heterogeneity of IoT infrastructures. To tackle these issues, this work introduces the Reconciliation-based IoT Application Management (RITAM) approach to manage IoT application deployments on IoT and edge infrastructures. It combines domain-specific modeling of IoT infrastructures and general-purpose modeling using TOSCA. To apply the RITAM approach, this work formalizes the Controller and Reconciler Pattern which replaces imperative management workflows with eventually consistent reconciliation. Moreover, the practical feasibility of RITAM is validated using a prototypical implementation.