Eliminating Training/Serving Skew with a Feature Store
A single declarative feature definition, compiled to both batch and online paths, ended an entire class of ML bugs for us.

## Two pipelines, one source of truth
Training/serving skew is insidious because it hides in plain sight. Here is how we unified feature computation behind one definition, enforced point-in-time correctness, and brought online lookups under 5ms.