Implementaction of lifecycle management infrastructure for machine learning models

This thesis addresses the implementation of Machine Learning and Operations (MLOps) within a Czech e-commerce company by leveraging the existing Development and Operations (DevOps) infrastructure and the integration of additional open-source components. The primary objective was to bridge the gap between machine learning (ML) model development and its deployment, which is a common challenge in the production use of ML. By leveraging the principles of DevOps, this work extends these methods to include machine learning features, implementing automated lifecycle management infrastructure for ML models. The implementation involved the analysis and adaptation of MLOps components suitable for the company’s technical infrastructure and operational workflows. A sequence-aware ML recommender system was developed to validate the implemented MLOps infrastructure. This system served as a proof of concept but also demonstrated practical applications in enhancing the customer experience and business operations. The integrated MLOps infrastructure was evaluated by the Google maturity model, resulting in the validation of the automated ML model lifecycle management infrastructure. Although the whole range of MLOps characteristics, such as Continuous Training, was not realized due to the absence of the model in production use, the work sets a foundation and outlook for future enhancements. The contribution of this thesis is the adaptation of existing DevOps infrastructure to support the MLOps workflows without introducing unnecessary complexity. This may serve as a starting point for other companies aiming to integrate MLOps. The implemented system poses limitations in direct transferability due to the tailored solution for the company context, but it provides a structured way for similar implementations. Future work will focus on the implementation of Continuous Training and more granular execution of the pipeline to enhance efficiency, reduce computational costs, and acceleration of deployment.