What is "production" Machine Learning?

In traditional software development, production systems are expected to meet various guarantees such as safety, traceability, observability, and scalability. These principles are similarly applicable to production machine learning (ML) systems, which involve the end-to-end training and inferencing pipeline. Safety in ML involves ensuring model inferences are within acceptable limits and do not harm users, achieved through unit testing and model simulations. Traceability is crucial, requiring exhaustive lineage tracking of all assets involved in model production, enabling auditability and debugging. Reproducibility is vital to ensure models can be recreated from raw data, allowing for experimentation and debugging. Automation is necessary for adapting models to changing data trends, requiring an automated end-to-end pipeline for retraining. Sematic, an open-source framework, provides production-grade guarantees by facilitating safety, traceability, reproducibility, and automation in ML pipelines, allowing users to build, execute, and manage complex ML processes effectively.