By extending DevOps principles to Data Sciences and Machine Learning, MLOps provides tools for the training, deployment and operation of AI models as ordinary software components that compose software architectures. A key issue of these processes is the continuous training of AI models to adapt them to observable changes in their input data coming from the production context (data and concept drifts). A no-code / low code solution seems adapted to the various user profiles that contribute to AI model development (data scientist, AI specialist or software engineer). Its design faces the challenges induced by the variety of components of MLOps processes. This project will study how domain engineering concepts (feature models, software product lines) could be used to reference commonalities in MLOps processes and document best practices for their use. The domain knowledge gathered and formalized this way will then be used to guide and assist the design of new MLOps pipelines. The approach will then be tooled by a Model Driven Engineering approach that will define a Domain Specific Language (DSL) that is altogether:
i) generic and extensible to cover the diversity of pipeline elements and target environments,
ii) abstract to be ready-to-use by non-expert users,
iii) yet allowing parameterization and fine-tuning by expert users,
iv) pivotal to obtain descriptions that will then be transformed into Platform-as-Code or Infrastructure-as-Code environments used for deployment. The project will focus on automatic and dynamic (re)-deployment of pipeline constituents that are hosted by multiple providers in order to optimize the efficiency, cost and environmental footprint of their operation. All proposals will be prototyped and validated through PoCs using the industrial partner’s platform (that provides cloud solutions) on real case studies provided to the industrial partner based on links with major French companies.
