gojek/feast

Features

• Feature Stores - Provides a centralized system for managing, storing, and serving curated data features for machine learning models.
• Feature Registries - Provides a central catalog for defining and discovering feature services used across multiple machine learning models.
• Historical Feature Retrieval - Extracts point-in-time correct feature sets from an offline store for model training and validation.
• Feature Definition Consistency - Stores and organizes feature definitions to ensure identical data definitions are used across training and production.
• On-Demand Feature Transformations - Provides a runtime computation engine that applies logic to raw input data immediately before serving it to models.
• Training Dataset Management - Provides unified tools for generating and validating point-in-time correct datasets used in model refinement.
• Training Dataset Processing - Generates point-in-time correct historical datasets for machine learning training while preventing data leakage.
• Feature Management - Provides a system for defining, serving, and managing machine learning feature sets and their versions.
• Feature Registries - Provides a central feature registry via CLI, SDK, and Web UI to track and manage model-centric feature services.
• Temporal Join Alignment - Aligns historical feature values with target labels using event timestamps to prevent temporal data leakage.
• Temporal Join Generators - Generates point-in-time correct historical datasets to prevent temporal data leakage during model training.
• Temporal Leakage Prevention - Generates feature sets that prevent data leakage by ensuring training data excludes future values relative to the target event.
• Low-Latency Data Retrieval - Implements high-speed retrieval of the latest feature values to power real-time model predictions in production.
• Real-Time Feature Computation - Computes feature values at request time or during write operations to derive real-time signals from input data.
• Real-time Feature Pipeline Orchestrators - Synchronizes processed feature data between offline storage and online serving layers for low-latency inference.
• Real-Time - Enables low-latency retrieval of the latest feature values to power live predictions in production.
• Vector Similarity Search - Provides indexing mechanisms for performing similarity queries on high-dimensional vector embeddings.
• Streaming Data Processing - Ingests and transforms data streams in real-time to push resulting features into online or offline stores.
• Multi-Source Data Integration - Connects to various databases, cloud warehouses, and streaming platforms to pull raw data for feature processing.
• Dual-Store Synchronization - Implements a synchronization process to mirror processed data across high-latency offline warehouses and low-latency online stores.
• Stream Ingestion - Implements a real-time pipeline that processes event streams and updates the online feature store.
• MLOps - Provides the underlying MLOps infrastructure to abstract feature retrieval and storage for training and inference consistency.
• Data Access Layers - Employs a unified data access layer to decouple feature retrieval from the underlying physical storage systems.
• Feature Access Abstractions - Decouples feature retrieval from underlying storage systems using a unified layer to improve model portability.
• Deep Learning Frameworks - Feature store for managing machine learning data.
• Machine Learning - Feature store for managing and accessing machine learning features.
• MLOps and Infrastructure - Feature store for managing and accessing ML features.

Star history