Implementing MLOps in the Enterprise

Chapter 6. Deployment of Models and AI Applications

Processing data, training, and validating models are precursors to the real thing: building and deploying an application that uses the data you generated and the model you have built to drive decisions and actions.

To deliver machine learning applications, start by building and registering the model(s) for use in the production application. Then, create an application pipeline that accepts events or data, prepares the required model features, infers results using one or more models, and drives actions. Finally, monitor the data, models, and applications to guarantee their availability and performance. In cases of problems or degraded model performance, drive corrective actions.

Many organizations still think of “serving a model” or creating a model endpoint. However, they need to pay more attention to the bigger picture of delivering an ML application as a whole instead of dividing the application delivery responsibility between data science and engineering teams. Ignoring the bigger picture will lead to significant functionality gaps, failures, unnecessary risks, and long delays.

Model Registry and Management

A model registry is a central repository for storing ML models and their metadata and managing the model lifecycle and versions. Once a model training process completes, it saves the model and its metadata in the registry. Then different functions (such as evaluation, testing, and optimization) extend the model metadata ...