AI-300 Practice Question 35: Feature Set Versions Point in Time
Microsoft Machine Learning Operations Engineer Associate (AI-300) · Test 4 · question 5 of 10
A churn model in production reads features from a managed feature store. A data scientist wants to add new window aggregates for a second model version and must not change what the first version sees in production. The training data for the new version is built from labeled events with event timestamps, and the team must not let feature values that were computed after each label's event time leak into the training rows. Which approach meets both requirements?
Show answer and explanation
Answer:
Managed feature store makes feature sets *"versioned and immutable"*, which *"allows you to independently manage the feature set lifecycle. You can deploy new model versions with different feature versions, and avoid disruption of the older model version."* For training data, *"feature retrieval supports point-in-time temporal joins (also known as time travel) to avoid data leakage"*, so each labeled event is matched with the feature values that were available at its event time. The built-in feature retrieval component or get_offline_features() generates that data (option B).
- A is wrong because feature sets are immutable, so an in-place edit is not possible, and the goal is not to change what the first version sees. - C is wrong because joining the latest value of each feature to each event uses values computed after the event, which is exactly the data leakage the team must avoid. - D is wrong because the online store holds current values for low-latency serving. It has no history to join at each event time, and offline retrieval is what generates training data.
Reference: What is managed feature store? - Azure Machine Learning