Chapter 9: Updating Production Models Using Amazon SageMaker Endpoint Production Variants
A deployed production model needs to be updated for a variety of reasons, such as to gain access to new training data, to experiment with a new algorithm and hyperparameters, or to model predictive performance deteriorating over time. Any time you update a model with a new version in production, there is a risk of the model becoming unavailable during the update and the model's quality being worse than the previous version. Even after careful evaluation in the development and QA environments, new models need additional testing, validation, and monitoring to make sure they work properly in production.
When deploying new versions of models into production, you should carefully consider reducing deployment risks and minimizing downtime for the model consumers. It is also important to proactively plan for an unsuccessful model update and roll back to a previous working model. Replacing an...
