Online and Offline Feature Serving Skew Quiz

Answers and explanations appear as you take the quiz.

1. What is training/serving skew?
   • When a model is trained on a GPU but served on a CPU
   • When a model is too large to fit in memory
   • When the features computed offline during training differ from those served online in production
   • When training takes longer than expected
2. When building a historical training row, which feature timestamps should a point-in-time join allow?
   • Values available at or before that row’s prediction time
   • Values recorded only after the outcome
   • The latest values regardless of historical time
   • Only values from the next calendar month
3. Which of the following is a classic cause of inflated offline accuracy that collapses in production?
   • Using too few training examples
   • Training on a faster machine than serving
   • Storing features in a database
   • Data leakage, where training uses information not available at prediction time
4. A logged live feature distribution differs from training. What should the feature-serving team infer first?
   • The serving code is certainly defective
   • There is a mismatch to investigate; it may be drift or serving skew
   • The feature has no predictive value
   • The model must be running on different hardware
5. Why is a cached online feature like 'orders in the last hour' a skew risk?
   • The cached value can go stale and no longer match the freshly computed value used in training
   • Caching always improves model accuracy
   • Caching makes the feature impossible to compute offline
   • Cached features cannot be numerical
6. Why share one feature definition across batch training and online serving?
   • To guarantee every cached value is fresh
   • To eliminate changes in incoming populations
   • To make timestamp checks unnecessary
   • To reduce differences caused by separate transformation implementations
7. A transaction model uses a chargeback flag recorded days after purchase. Why is its offline feature join invalid?
   • The flag is categorical rather than numerical
   • The flag was unavailable when the purchase prediction was made
   • Chargeback flags cannot be stored historically
   • Only GPU models may use transaction features
8. Training recomputes a one-hour order count, but serving reads a ten-minute-old cached count. Which check targets this skew?
   • Increase the number of model parameters
   • Remove all timestamps from the feature store
   • Compare counts for matching entities and times while recording cache freshness
   • Switch both paths to a different programming language