Browsing by Author "Yumo, Luo"

Machine Learning Operations (MLOps), derived from DevOps, aims to unify the development, deployment, and maintenance of machine learning (ML) models. Continuous training (CT) automatically retrains ML models, and continuous deployment (CD) automatically deploys the retrained models to production. Therefore, they are essential for maintaining ML model performance in dynamic production environments. The existing proprietary solutions suffer from drawbacks such as a lack of transparency and potential vendor lock-in. Additionally, current MLOps pipelines built using open-source tools still lack flexible CT and CD for ML models. This study proposes a cloud-agnostic and open-source MLOps pipeline that enables users to retrain and redeploy their ML models flexibly. We applied the Design Science methodology, consisting of identifying the problem, defining the solution objectives, and implementing, demonstrating, and evaluating the solution. The resulting solution is an MLOps pipeline called CTCD-e MLOps pipeline. We formed a conceptual model of the needed functionalities of our MLOps pipeline and implemented the pipeline using only open-source tools. The CTCD-e MLOps pipeline runs atop Kubernetes. It can autonomously adapt ML models to dynamic production data by automatically starting retraining ML models when their performance degrades. It can also automatically A/B test the performance of the retrained models in production and fully deploys them only when they outperform their predecessors. Our demonstration and evaluation of the CTCD-e MLOps pipeline show that it is cloud-agnostic and can also be installed in on-premises environments. Additionally, the CTCD-e MLOps pipeline enables its users to flexibly configure model retraining and redeployment as well as production A/B test of the retrained models based on various requirements.