Published Paper Details:

Using Machine Learning in CI/CD pipelines brings about specific challenges since models depend on data (data can vary), they evolve over time during the training/exploitation stage (initially a single model is produced for deployment/testing), and the model may require continued training and/or retraining. CI/CD pipelines as they exist today are meant for deterministic code (e.g., non-ML libraries) and do not take into consideration additional important ML system metrics such as data drift and model performance. In this paper, we summarize and highlight the major issues for ML CI/CD implementation, we describe some concrete architecture patterns to deal with identified issues, and give a case study using GitHub Actions and Azure Machine Learning. Results on evaluation show that MLOps based CI/CD has increased ability to release more rapidly, the full pipeline is more reliable, and operationalizing bad models can be avoided.