SATURN 2020

Most of the time, machine learning (ML) is strongly viewed from a data science perspective. This means, you can find tons of information on algorithms and the treatment of data. However, what it actually means to architect systems, in which ML plays a role, is rather rarely found. Our focus is on the engineering of typically large systems, which are, to some degree, basing their functionality on machine learning. As such systems often serve in production large amounts of users, the fulfillment of quality attributes is critical and needs consideration in architecture design.

In this talk, we systematically decompose in the language of software architects what it means to build a system based on machine learning. We outline an architectural design space and discuss central architecture decisions an architect has to make when designing a system based on ML.

• This includes a perspective on both, the development time, and the runtime.
• We show how a system can be decomposed and how machine learning components look like and behave in the context of an overall system.
• Machine learning is fundamentally depending on data: ; Thus, the data aspect is central in our architectural considerations.
• As neural networks are very widespread nowadays for the realization of ML-based systems, we take a closer look at their architectural implications.
• We include a perspective on the activities around data collection, preparation, model selection and training, and model inference.
• We discuss deployment options for model training and model inference.
• We discuss different types of technologies available for machine learning, from as-a-service APIs over pre-trained models down to pure libraries requiring to construct and to the train the full model.

With this overview, architects will get the big picture of designing ML-based systems and have a much better position to bridge the gaps between data scientists, data engineers, and software developers and architects.