midMLOps
What is training-serving skew?
Updated May 17, 2026
Short answer
Training-serving skew occurs when feature computation differs between training and production.
Deep explanation
This inconsistency leads to degraded model performance in production. It often happens due to different preprocessing pipelines or missing feature parity. MLOps solves this using feature stores and shared transformation pipelines.
Real-world example
A model trained with normalized features fails because production inputs are raw values.
Common mistakes
- Rewriting feature logic separately in training and serving systems.
Follow-up questions
- How to detect skew?
- How to prevent it?