PRODUCTION SCENARIO
A model trained on transaction amounts in dollars is deployed behind a real-time endpoint. The upstream service sends amounts in cents, and predictions are systematically wrong although offline validation was strong. Three other models consume the same amount feature, and the fix must cover all of them.
Which change repairs the cause across every consumer?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Apply one shared transformation on the training path and the serving path
This is training-serving skew: the serving path presents a different representation than the model was fit on. One shared transformation restores the input contract everywhere the feature is read. Converting inside one inference container leaves the other three models broken, retraining on cents breaks any caller still sending dollars, and a range constraint only reports the mismatch.
Official Source: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html ↗