Feature store

As explained in Machine Learning Platform Engineering

A feature store is a system that centrally stores, manages, and serves curated machine-learning features for training and inference.

A feature store makes curated features available for both model training and inference. Instead of each team rebuilding transformations or maintaining its own retrieval queries, a shared feature registry catalogs what is available. Using the same managed features across training and serving helps prevent training-serving skew. section 1.3.3 section 4.1.3

Feast fits after feature computation: a pipeline produces transformed feature data in an offline store, such as a MinIO bucket, while Feast registers feature definitions and provides retrieval methods. Keeping feature generation separate from modeling makes the resulting features easier to share and reuse. section 4.2

Figure 4.10 Feast feature store design involves a feature pipeline populating the offline store and periodically Feast materializes the offline features to the online store. The feature registry holds feature definitions along with online and offline store information. The Feast SDK provides methods to retrieve features from online and offline stores, which can be used for training and inference purposes.

The diagram separates the feature pipeline, offline store, online store, registry, and Feast SDK. Follow the path from computed features to historical retrieval for training or low-latency retrieval for inference. section 4.2

Retrieve features as they existed at the time

Suppose a user has feature records dated January 7, May 21, and December 4. For a May 22 retrieval, the appropriate record is May 21—not the later December record. Feast can produce point-in-time-correct feature sets without a separate SQL join or retrieval script. section 4.2

Choose retrieval according to the workload:

• Use an offline store for historical features during training or batch inference. section 1.3.3 section 4.3
• Materialize offline features into an online store, such as Redis, when real-time inference needs low-latency retrieval. section 4.2

A feature store complements rather than replaces the rest of the ML platform: MLflow tracks experiments and registers models, while Feast manages their features. The Feast UI can show registered features, entities, and data sources. Updating features and retraining models still require orchestration; registration alone does not automate those jobs. section 4.0 section 4.2.4

How Machine Learning Platform Engineering covers it

How deep it goes

Explains Feast’s registry, offline and online stores, materialization, and point-in-time retrieval; leaves retraining and feature-update automation to pipelines.

Examples

Uses timestamped user features for point-in-time retrieval and a tabular movie recommender to motivate feature-store tooling.

What it assumes

Assumes familiarity with model training and feature engineering, while introducing why their outputs need registration and consistent retrieval.

What you can do after

Apply it via Feast to register and share features, retrieve historical feature sets, and serve features from an online store.

How it connects

Connects feature engineering and training-serving skew with MLflow model registration, online inference, and Kubeflow pipeline automation.

liveBookmanning.com

Read Machine Learning Platform Engineering on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  LLMs in ProductionA feature store is a system or abstraction for managing, documenting, versioning, governing, and serving machine-learning features for training and production.
• 

  liveBookmanning.com
  Machine Learning System DesignAn infrastructure system for storing, managing, and serving machine-learning features.
• 

  liveBookmanning.com
  Delivering Machine Learning Projects: From design to deploymentA feature store is a system for storing, managing, and serving machine learning features for model development and inference.
• 

  liveBookmanning.com
  Financial AI in PracticeA centralized repository for storing, managing, and serving validated features for machine-learning workflows.

Compare how 5 books explain feature store →

Related concepts

• inputuses · An input is a value, information, or signal supplied to a system, computation, or interface for processing or to influence its behavior.
• training-serving skewaddresses · Training-serving skew is a mismatch between the data, feature representations, or computations used during model training and those used to generate predictions.
• Apache Airflowcontrasts with · Apache Airflow is an open-source platform for authoring, scheduling, and monitoring programmatically defined workflows, commonly represented as directed acyclic graphs.
• machine learninguses · Machine learning is a field of computational methods in which systems learn patterns or relationships from data to make predictions, decisions, or other outputs.
• independent variableuses · An independent variable is an input or explanatory attribute whose values are used to model, explain, or predict variation in an outcome.
• source of truthaddresses · A source of truth is the authoritative source or representation from which the accepted value or state of information is determined.

This explainer was generated by Manning's knowledge graph from the text of Machine Learning Platform Engineering. The section links point to the parts of the book it draws on.