hard20 pts · first answer countstraining-serving-skewfeature-store

Great offline, bad in production

A model scores well on a held-out test set, but after deployment its predictions are clearly worse. Investigation shows "avg_spend_30d" is computed in SQL for training but in a separate Java service at serving time. What is the best long-term fix?

A.Compute features once with shared logic (e.g. a feature store) for both training and serving, with point-in-time correct training dataB.Retrain the model every hour so it adapts to the Java valuesC.Add more layers to the model so it can learn around the differencesD.Remove the feature and accept lower accuracy

Discussion 0

Join free to post in the discussion.

Great offline, bad in production

Discussion 0

Related questions