Machine-learning-enabled systems are becoming increasingly common in different industries. Due to the impact of uncertainty and the pronounced role of data, ensuring the quality of such systems requires consideration of several unique characteristics in addition to traditional ones. This range of quality attributes can be achieved by the implementation of specific architectural tactics. Such architectural decisions affect the further functioning of the system and its compliance with business goals. Architectural decisions have to be made with attention to possible quality trade-offs to prevent the cost of mitigating unintended side effects. A related work analysis revealed the need for a thorough study of existing architectural decisions and their impact on various quality attributes in the context of machine-learning-enabled systems. In this paper, to address this goal, we present comprehensive research on the quality of such systems, architectural tactics, and their possible quality consequences. Based on a systematic literature review of 206 primary sources, we identified 11 common quality attributes, and 16 relevant architectural tactics together along with 85 potential quality trade-offs. Our results systematize existing research in building architectures of ML-enabled systems. They can be used by software architects and researchers at the system design stage to estimate the possible consequences of decisions made. • A Common Quality Model for ML-enabled systems. • A List of Architectural Tactics to Achieve Identified Quality Attributes. • A Broad Analysis of Quality Trade-offs.

Machine Learning (ML) - enabled systems capture new frontiers of industrial use. The development of such systems is becoming a priority course for many vendors due to the unique capabilities of Artificial Intelligence (AI) techniques. The current trend today is to integrate ML functionality into complex systems as architectural components. There are a lot of relevant challenges associated with this strategy in terms of the overall system architecture and in the context of development workflow (MLOps). The probabilistic nature, crucial dependency on data, and work in an environment of high uncertainty do not allow software engineers to apply traditional software development methodologies. As a result, there is a community request to systematize the most relevant experience in building software architectures with ML components, to create new approaches to organizing the process of developing ML-enabled systems, and to build new models for assessing the system quality. Our research contributes to all mentioned directions and aims to create a methodology for the efficient implementation of ML-enabled software and AI components. The results of the research can be used in the design and development in industrial settings, as well as a basis for further studies in the research field, which is of both practical and scientific value.