163
train-serve skew
The feature in training is not the feature in production. Same name, different code.
What is train-serve skew?
Skew is a mismatch between how a feature is computed offline and online: a join that sees the future offline, a null treated differently, a Python function the server reimplemented wrong. Accuracy collapses only after deploy.
Why does train-serve skew matter when vibe coding?
Models notebook a feature, then “port it” to the API from memory. Name one implementation shared by both paths.
How do you do train-serve skew?
One function or one feature definition used in training and serving. Log the served features. Diff them against the training row for the same key.
How do you ask a model for train-serve skew?
Prevent train-serve skew for (feature). One implementation used both offline and online. Log served feature values. Do not reimplement the notebook in the API by hand.
What goes wrong with train-serve skew?
Sharing a function that reads “now” so offline replay and online calls still differ. Pass the timestamp in.