Training features are computed one way, but the production API computes them differently. What risk should be investigated?

1. Only a harmless display difference
2. A database DNS propagation delay
3. Training-serving skew
4. A guarantee of better generalization