Why your ML team needs a feature store (and what to build first)

Ask any ML team what their biggest time sink is and the answer is almost never "training models." It's data preparation — the same feature transformations written over and over, in slightly different ways, for different models.

The three problems a feature store solves

Duplication: if your churn model and recommendation model both need "days since last purchase," they should both read from the same transformation. Training-serving skew: if training computes features differently from serving, your model behaves differently in production. Discoverability: without a store, the answer to "what features exist?" is "ask someone."

What to build first

Don't start with the infrastructure. Start with the registry. A simple catalogue of what features exist, how they're defined, and which models use them is valuable even before you have a formal store.

Once you have a registry, add versioning. Features change, and models depend on specific versions. If you don't version your features, you can't safely iterate on them without risk of breaking downstream consumers.

Infrastructure comes last. By the time you actually build or adopt a feature store platform, you'll have a clear picture of what you need it to do — and you'll avoid over-engineering for a problem you don't have yet.