Training data was normalized with min–max scaling in AWS Glue DataBrew. Production inference data must be normalized exactly the same way before being sent to the model. Which approach satisfies this requirement?

The correct approach is to retain the min-max normalization parameters (minimum and maximum values for each feature) computed from the training data and apply those exact same values to normalize the production inference data. This ensures consistency between the data the model was trained on and the data it receives for predictions. If different normalization parameters are used for inference, the model will receive data in a different distribution than it expects, leading to inaccurate predictions. Applying normalization statistics from a public dataset is incorrect because those statistics will not match the specific distribution of your training data. Computing new min-max ranges from production samples or individual samples is incorrect because it introduces data leakage and inconsistency with the training data's scaling.