AI glossary · Advanced · MLOps & LLMOps · Data engineering
Feature store
A feature store is a central repository where the input values (features) used by machine learning models are created and stored. It lets the same values be retrieved in exactly the same way both during training and when making predictions in a live service.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Data, ops & platforms
- Fields
- MLOps & LLMOps, Data engineering
- Added
- 2026-10-08
How it works
First, you define a pipeline that processes raw data (purchase records, click logs, etc.) into features such as 'number of purchases in the last 7 days.' These features are stored in two places. One is the offline store, which keeps a long history of past records and is used to build training data for models. The other is the online store, which can quickly return only the latest values and is used for real-time predictions while the service is running. Information such as each feature's name, definition and creator is registered as well, so other teams can find and use it.
Why it matters
The most common problem is that the way a value is computed during training and the way it is computed in the service drift apart, quietly degrading model performance. A feature store reduces this gap by having the definition written once and used on both sides. It also cuts down on duplication, where several teams each rebuild the same feature. It can also build training data using only the values that were knowable at a specific point in the past, which prevents future information from slipping in by mistake.
Things to keep in mind
Feature stores are mostly used in traditional machine learning that handles tabular data, such as recommendation, anomaly detection and demand forecasting. In LLM services, a vector DB for document search sometimes plays a partly similar role, but the two serve different purposes. For small organizations with only a few models, the burden of adopting and operating one can outweigh the benefits. It is often adopted as one part of a set of MLOps tools.
Example
Take a shopping app's recommender system as an example: features such as 'product categories this user viewed recently' or 'amount spent last week' are registered in the feature store. Data scientists train the recommendation model on past values from the offline store, and when a user opens a screen in the app, the server immediately pulls the latest values of the same features from the online store and feeds them into the model to produce a recommendation list. A fraud detection model can use the same features as well.