What is training-serving skew?

Training-serving skew is when features are computed differently during training than at serving time — different code paths, data sources, or timing — so the model sees inconsistent inputs and underperforms in production despite good offline metrics. Prevent it by sharing the same feature-transformation code between training and serving (or a feature store), logging serving features to compare against training, and testing the full pipeline end to end. It's a leading cause of 'great in testing, bad in production'.