During training, the version of packages such as scikit-learn is not defined, so the latest version is picked up:

During scoring, the version of most (but not all) packages is set:

This led to binary incompatibility of the saved model when a new version of scikit-learn was released: #127, #126

For IaC best practices, we should use fixed versions for all packages in all environments.

Ideally we should also have a mechanism to define in a single place the names and versions of packages needed in training and scoring, to avoid redundancy and the risk of divergence.