As the level of digitization in industrial environments increases, companies are striving to improve efficiency and resilience to unplanned disruptions through the development of machine learning (ML)-based applications. Still, sustainable deployment and operation beyond proofs-of-concept is a challenging and resource-intensive task in dynamic enviroments such as industry 4.0, often impeding practical adoption in the long-term and thus sustainable ML product development. In this work, we systematically identify these challenges based on the CRISP-ML process model phases by applying a design science research approach. To this end, we conducted 15 interviews with data science practitioners in industry 4.0. Following a qualitative content analysis, design requirements and design principles for the development and sustainable long-term deployment of ML systems are derived to address identified challenges such as robustness to, and management of data drift caused by time-dependencies and machine/product differences, missing metadata, interfaces to other IT systems, expectation management, and MLOps guidelines.