PRODUCTION SCENARIO
A retailer trained a churn model on a days-since-last-purchase feature that a nightly Apache Spark job computes. The web team reimplemented the same feature in application code for inference. Production predictions now differ from the offline results.
Which solution will address the root cause with the LEAST operational overhead?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Compute the feature once and serve it from Amazon SageMaker Feature Store.
Two separate implementations of the same feature produce training-serving skew. Feature Store lets one feature processor ingest values that training reads from the offline store and inference reads from the online store, so both paths see identical features with no service to build. A self-built shared service reaches the same end state but the retailer operates it. Retraining teaches the model the drifted values, and a drift alert reports the gap without closing it.
Official Source: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html ↗