feature-store-consistency

name	feature-store-consistency
description	Prevent training-serving skew: compute features once via a feature store or shared transformation so training and serving use identical logic, with point-in-time correctness for temporal features and no leakage of future data.

Feature Store / Train-Serve Consistency

The #1 production ML bug

'Great offline, bad online' is almost always training-serving skew — features computed differently at serving time.

Fix

Compute features once (feature store / shared transform); training and serving read the same logic.

Point-in-time correctness

For temporal features, join as-of the event time — no future leakage into training. Otherwise the offline metric is a lie.