Architectural Run-time Models to Facilitate Quality-aware DevOps

Building software applications by composing cloud services promises many benefits such as flexibility and scalability. However, it leads to major challenges like increased complexity, fragility and changes during operations that are unknown in development phase. These rapidly changing applications require communication and collaboration between software developers and operators as well as strong integration of building, evolving and adaptation activities. DevOps is a set of practices to enable software developers and operators to work more closely and thus reduce the time between changing a system and putting the change into production, while ensuring high quality [6]. The software application architecture is a central artifact at the interface between development and operations. The phase-spanning consideration of software architecture is essential in DevOps practices. New architectural styles such as microservices are proclaimed to satisfy requirements like scalability, deployability and continuous delivery. By simply introducing new architectural styles, however, current problems in the collaboration and communication among stakeholders of the development and operation phases are not solved. Existing architectural models used in the development phase differ from those used in the operation phase. Three differences are highlighted and described hereafter. These differences result in limited reuse of development models during operations and limited phasespanning consideration of the software architecture. Abstraction levels differ in architectural models from development and operations. Architectural models especially in early phases of development often adhere to the component-based paradigm to keep track of the architecture and structure the application. Instances of the Palladio Component Model (PCM) [9] are examples of architectural models used in development. During operations architectural models are applied to reflect the executed application, e.g. its current deployment or usage intensity. Architectural models used in operations are closer to an implementation level of abstraction (e.g., reflect method calls and class signatures) as they result from monitoring data related to source code artifacts. It is hard to reproduce component models from monitoring data as