How feast-dev/feast works
Feast is a feature store for ML: teams define and register features, retrieve point-in-time-correct historical data for training, materialize feature values to an online store, and retrieve them for inference. The Python SDK and storage/compute integrations provide the main workflow; serving APIs, the web UI, and Kubernetes operator support deployment and management. Workflow links reflect the README; detailed runtime wiring was not sampled, so uncertain implementation-level connections are omitted.
Main entry points: feature_store.py FeatureStore SDK · __init__.py Online stores · retrieval_task.py Historical retrieval · materialization_job.py Materialization
Components 24
SDK and features
- FeatureStore SDK
sdk/python/feast/feature_store.py - Feature definitions
sdk/python/feast/feature_view.py - Feature services
sdk/python/feast/feature_service.py
Compute and storage
- Feature registry
sdk/python/feast/infra/infra_object.py - Data sources
sdk/python/feast/data_source.py - Offline stores
sdk/python/feast/infra/offline_stores/offline_store.py - Historical retrieval
sdk/python/feast/infra/common/retrieval_task.py - Compute engines
sdk/python/feast/infra/compute_engines/base.py - Materialization
sdk/python/feast/infra/common/materialization_job.py - Online stores
sdk/python/feast/infra/online_stores/__init__.py - Vector store support
sdk/python/feast/infra/online_stores/faiss_online_store.py
Serving and clients
- Feature server
sdk/python/feast/feature_server.py - Online retrieval API
sdk/python/feast/embedded_go/online_features_service.py - Go serving runtime
go/internal/feast/onlineserving/serving.go - Java serving runtime
java/serving/src/main/java/feast/serving/service/OnlineServingServiceV2.java
Management and deployment
- Registry REST API
sdk/python/feast/api/registry/rest/rest_registry_server.py - Feast CLI
sdk/python/feast/cli/cli.py - Web UI
ui/src/queries/useResourceQuery.ts - Kubernetes operator
infra/feast-operator/internal/controller/featurestore_controller.go
Other components
- ML platform team
- Data scientist
- Model application
- Training dataset
- Inference features
Connections 26
- ML platform team → Feature definitions: defines
- Feature definitions → FeatureStore SDK: configures
- ML platform team → Feast CLI: runs apply
- Feast CLI → FeatureStore SDK: invokes
- FeatureStore SDK → Feature registry: registers definitions
- Data scientist → FeatureStore SDK: requests history
- FeatureStore SDK → Historical retrieval: requests retrieval
- Historical retrieval → Feature registry: reads definitions
- Historical retrieval → Offline stores: queries history
- Offline stores → Data sources: reads data
- Historical retrieval → Training dataset: produces
- FeatureStore SDK → Materialization: starts materialization
- Materialization → Compute engines: uses engine
- Materialization → Offline stores: reads feature data
- Materialization → Online stores: writes features
- Model application → FeatureStore SDK: requests online features
- FeatureStore SDK → Online retrieval API: delegates retrieval
- Online retrieval API → Online stores: reads features
- Feature server → Online stores: serves features
- Go serving runtime → Online stores: retrieves features
- Java serving runtime → Online stores: retrieves features
- Online retrieval API → Inference features: returns features
- Feature services → Feature definitions: projects features
- Web UI → Registry REST API: fetches resources
- ML platform team → Kubernetes operator: manages deployment
- Vector store support → Online stores: supports vector retrieval