# https://github.com/feast-dev/feast Project Manual Generated at: 2026-06-20 19:48:24 UTC ## Table of Contents - [Feast Overview, Architecture, and Core Concepts](#page-1) - [Data Stores, Materialization, and Feature Retrieval](#page-2) - [Feature Servers, Compute Engines, and Transformation Pipeline](#page-3) - [Deployment, Kubernetes Operator, Web UI, and Operations](#page-4) ## Feast Overview, Architecture, and Core Concepts ### Related Pages Related topics: [Data Stores, Materialization, and Feature Retrieval](#page-2), [Feature Servers, Compute Engines, and Transformation Pipeline](#page-3), [Deployment, Kubernetes Operator, Web UI, and Operations](#page-4)
Related Source Files The following source files were used to generate this page: - [README.md](https://github.com/feast-dev/feast/blob/main/README.md) - [infra/feast-operator/README.md](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md) - [sdk/python/feast/templates/local/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/local/README.md) - [sdk/python/feast/templates/gcp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/gcp/README.md) - [sdk/python/feast/templates/milvus/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/milvus/README.md) - [sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md) - [java/serving/README.md](https://github.com/feast-dev/feast/blob/main/java/serving/README.md) - [ui/README.md](https://github.com/feast-dev/feast/blob/main/ui/README.md) - [infra/charts/feast/charts/feature-server/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/feature-server/README.md) - [infra/charts/feast/charts/transformation-service/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md) - [examples/rag/README.md](https://github.com/feast-dev/feast/blob/main/examples/rag/README.md) - [examples/rag-docling/README.md](https://github.com/feast-dev/feast/blob/main/examples/rag-docling/README.md) - [examples/rag-retriever/README.md](https://github.com/feast-dev/feast/blob/main/examples/rag-retriever/README.md) - [examples/credit-risk-end-to-end/README.md](https://github.com/feast-dev/feast/blob/main/examples/credit-risk-end-to-end/README.md)
# Feast Overview, Architecture, and Core Concepts ## 1. Purpose and Scope Feast (**Fea**ture **St**ore) is an open source feature store for machine learning. According to the project README, Feast is positioned as "the fastest path to manage existing infrastructure to productionize analytic data for model training and online inference" ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). It is designed to serve ML platform teams who must: - **Make features consistently available for training and serving** by managing an *offline store* (historical, scale-out), a low-latency *online store* (real-time prediction), and a battle-tested *feature server* ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). - **Avoid data leakage** by generating point-in-time correct feature sets so future feature values never leak into training data ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). - **Decouple ML from data infrastructure** by exposing a single data access layer that abstracts feature storage from feature retrieval ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). The minimal Feast deployment diagram lives at `docs/assets/feast_marchitecture.png` ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). Feast is a community project still under active development, with releases published regularly (v0.55.0 through v0.64.0 are tracked in the release history). ## 2. High-Level Architecture Feast is organized as a multi-language ecosystem built around a declarative feature repository. The Python SDK is the primary user-facing surface; Java provides the high-throughput serving layer; the Go-based Kubernetes operator manages deployments. ```mermaid flowchart LR A[Feature Repository
feature_store.yaml + Python definitions] --> B[Feast CLI / SDK
feast apply, materialize] B --> C[Registry
file or SQL] B --> D[Offline Store
BigQuery, Snowflake, Redshift, Parquet, File] B --> E[Online Store
Redis, SQLite, DynamoDB, etc.] D -->|materialize| E E --> F[Feature Server
Python feature_server.py or Java serving] F --> G[Online Inference
get_online_features] F --> H[Transformation Service
on-demand features] K[Feast Operator
FeatureStore CR] --> F L[Feast Web UI] --> C ``` Key runtime pieces: - **Feature Repository** — A version-controlled directory containing `feature_store.yaml` plus Python files that declare entities, feature views, and feature services ([sdk/python/feast/templates/local/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/local/README.md)). - **Registry** — A file or SQL-backed catalog that records all defined objects. Templates document configuring a remote registry like `s3://my-bucket/feast/registry.pb` for production ([sdk/python/feast/templates/gcp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/gcp/README.md)). - **Feature Server** — Exposes HTTP/gRPC endpoints for online retrieval. The Java implementation lives in `java/serving/` and is shipped as an executable jar ([java/serving/README.md](https://github.com/feast-dev/feast/blob/main/java/serving/README.md)). The Python equivalent is started via `feast serve --host 0.0.0.0 --port 6566` ([sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md)). - **Transformation Service** — A separate Helm sub-chart that computes on-demand features at request time ([infra/charts/feast/charts/transformation-service/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md)). - **Feast Operator** — A Kubernetes operator that deploys/manages Feast via the `FeatureStore` Custom Resource ([infra/feast-operator/README.md](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md)). - **Feast Web UI** — An experimental React UI launched via `feast ui` or imported as a module ([ui/README.md](https://github.com/feast-dev/feast/blob/main/ui/README.md)). ## 3. Core Concepts ### 3.1 Feature Repository and Templates A feature repository is bootstrapped with `feast init` and a template. Templates shipped in `sdk/python/feast/templates/` include `local`, `gcp`, `aws`, `snowflake`, `milvus`, and `pytorch_nlp` ([sdk/python/feast/templates/local/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/local/README.md), [sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md)). Each template scaffolds a `feature_store.yaml`, demo data under `data/`, and a `test_workflow.py` that exercises the main Feast commands ([sdk/python/feast/templates/local/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/local/README.md)). The default `local` template configures SQLite for the online store, a file registry at `data/registry.db`, and a file offline store ([sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md)). ### 3.2 Entities, Feature Views, and Feature Services Feature definitions live in Python files such as `example_repo.py` or `feature_definitions.py` ([examples/rag/README.md](https://github.com/feast-dev/feast/blob/main/examples/rag/README.md)). Declarative objects include: - **Entities** — join keys (e.g., `customer_id`, `driver_id`) used to look up features. - **Feature Views** — groups of features sourced from a single data source, including `BatchFeatureView`, `StreamFeatureView`, and on-demand variants. Production configurations in the GCP/AWS templates wire these to Redshift, BigQuery, or Snowflake ([sdk/python/feast/templates/gcp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/gcp/README.md)). - **Feature Services** — named bundles of features consumed by a specific model version. The PyTorch NLP template defines `sentiment_analysis_v1`, `sentiment_analysis_v2`, and `sentiment_training_features` ([sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md)). ### 3.3 Data Sources and Stores The README enumerates supported data sources: Snowflake, Redshift, BigQuery, Parquet file, Azure Synapse + Azure SQL (contrib), Hive (community), Postgres (contrib), Spark (contrib), and Couchbase (contrib) ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). Online stores include Redis (default for production) and SQLite (default for demos) ([sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md)). The Milvus template demonstrates vector-database integration for RAG workloads ([sdk/python/feast/templates/milvus/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/milvus/README.md)). ### 3.4 Serving and Retrieval Two retrieval paths are documented: - **Online retrieval** — `get_online_features` queries the online store via the feature server ([sdk/python/feast/templates/gcp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/gcp/README.md)). - **Offline retrieval** — historical feature joins for training, produced from the offline store with point-in-time correctness ([README.md](https://github.com/feast-dev/feast/blob/main/README.md)). A `DocEmbedder` helper class automates the document-to-embeddings pipeline (chunking, embedding, `FeatureView` creation, online write) for RAG ingestion ([examples/rag-retriever/README.md](https://github.com/feast-dev/feast/blob/main/examples/rag-retriever/README.md)). The Docling example extends this by transforming PDFs into Parquet chunks suitable for embedding ([examples/rag-docling/README.md](https://github.com/feast-dev/feast/blob/main/examples/rag-docling/README.md)). ## 4. Deployment Components | Component | Path | Purpose | |---|---|---| | Feature Server (Java) | [java/serving/](https://github.com/feast-dev/feast/blob/main/java/serving/README.md) | High-throughput online retrieval server | | Feature Server (Python) | `feast serve` CLI ([sdk/python/feast/templates/pytorch_nlp/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/pytorch_nlp/README.md)) | Lightweight online retrieval | | Transformation Service | [infra/charts/feast/charts/transformation-service/](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md) | On-demand feature computation | | Helm chart | [infra/charts/feast/](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/feature-server/README.md) | Kubernetes deployment of feature server | | Feast Operator | [infra/feast-operator/](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md) | Manages `FeatureStore` CRs; supports RBAC/OIDC, TLS, batch jobs, registries | | Web UI | [ui/](https://github.com/feast-dev/feast/blob/main/ui/README.md) | Browses the registry; embeddable as `@feast-dev/feast-ui` | ## 5. Known Issues and Community Notes Several community-reported issues and release notes are relevant when adopting Feast: - **UI regression between v0.63.0 and v0.64.0** — the Feast Web UI reportedly breaks in the demo after upgrading ([issue #6539](https://github.com/feast-dev/feast/issues/6539)). - **Missing zoned datetime feature type** — Feast does not yet natively store timezone information alongside timestamps; users currently split zone info into a separate feature ([issue #6536](https://github.com/feast-dev/feast/issues/6536)). - **`get_online_features` drops join keys on remote stores** — when calling the SDK against remote feature servers, only the first entity key of a multi-join-key entity row is honored ([issue #6488](https://github.com/feast-dev/feast/issues/6488)). - **v0.64.0 fixes** — added `async_supported` to `RedisOnlineStore` and supplied missing `feast init` templates for the operator CRD ([v0.64.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.64.0)). - **v0.55.0–v0.63.0 fixes** — improvements to PostgreSQL subquery aliasing, Trino-to-Feast type mapping, `unix_timestamp_val` serialization, transformation persistence, and project-filtered `apply_data_source`/`delete_data_source` ([release notes](https://github.com/feast-dev/feast/releases)). ## 6. See Also - [Feast Documentation](https://docs.feast.dev/) - [Quickstart](https://docs.feast.dev/getting-started/quickstart) - [Examples Index](https://github.com/feast-dev/feast/tree/master/examples) - [Feast Operator Guides](https://docs.feast.dev/how-to-guides/feast-operator) - [Java Serving Developer Guide](https://github.com/feast-dev/feast/blob/main/java/serving/README.md) - [RAG Examples](https://github.com/feast-dev/feast/blob/main/examples/rag/README.md) --- ## Data Stores, Materialization, and Feature Retrieval ### Related Pages Related topics: [Feast Overview, Architecture, and Core Concepts](#page-1), [Feature Servers, Compute Engines, and Transformation Pipeline](#page-3)
Related Source Files The following source files were used to generate this page: - [sdk/python/feast/infra/online_stores/online_store.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/online_store.py) - [sdk/python/feast/infra/online_stores/redis.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/redis.py) - [sdk/python/feast/infra/online_stores/remote.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/remote.py) - [sdk/python/feast/infra/online_stores/hybrid_online_store/hybrid_online_store.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/hybrid_online_store/hybrid_online_store.py) - [sdk/python/feast/infra/online_stores/milvus_online_store/milvus.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/milvus_online_store/milvus.py) - [sdk/python/feast/infra/offline_stores/offline_store.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/offline_stores/offline_store.py) - [go/internal/feast/server/logging/featureserviceschema.go](https://github.com/feast-dev/feast/blob/main/go/internal/feast/server/logging/featureserviceschema.go) - [README.md](https://github.com/feast-dev/feast/blob/main/README.md)
# Data Stores, Materialization, and Feature Retrieval ## Purpose and Scope Feast is an open source feature store that bridges offline historical data and low-latency online serving. The system is organized around three primary storage and access components: an **offline store** for scalable historical feature generation and point-in-time joins, an **online store** for real-time feature retrieval, and a **materialization** process that continuously moves feature values from the offline store into the online store. The repository exposes these as pluggable backends so teams can pick the infrastructure that matches their scale and latency requirements. Source: [README.md](https://github.com/feast-dev/feast/blob/main/README.md). This page documents the online and offline store abstractions, the materialization flow, and the feature retrieval API surface, including the join-key handling that is central to both online reads and on-demand feature views. ## Online Stores The `OnlineStore` protocol defines the contract every backend must implement. It is defined in [sdk/python/feast/infra/online_stores/online_store.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/online_store.py) and is the surface through which feature servers read pre-computed feature values. Backends included in the repository include: | Backend | Source | Typical Use | |---|---|---| | Redis | [sdk/python/feast/infra/online_stores/redis.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/redis.py) | Default low-latency key-value store | | Remote | [sdk/python/feast/infra/online_stores/remote.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/remote.py) | Forwards requests to a remote Feast server | | Hybrid | [sdk/python/feast/infra/online_stores/hybrid_online_store/hybrid_online_store.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/hybrid_online_store/hybrid_online_store.py) | Combines a fast local cache with a remote source | | Milvus | [sdk/python/feast/infra/online_stores/milvus_online_store/milvus.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/online_stores/milvus_online_store/milvus.py) | Vector similarity search for RAG and embedding features | Version notes from the community: v0.64.0 added an `async_supported` property to `RedisOnlineStore` to better signal whether the backend can be used in async serving paths ([v0.64.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.64.0)), and v0.59.0 introduced the hybrid online store as a new official backend ([v0.59.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.59.0)). ### Join-Key Handling The online read API accepts a list of entity keys, each potentially composed of multiple join keys, and must return a row per entity key. The Go feature server encodes the resolved join keys, their types, and the request shape in `getFeatureServiceSchema` for logging and observability. Source: [go/internal/feast/server/logging/featureserviceschema.go:21-65](https://github.com/feast-dev/feast/blob/main/go/internal/feast/server/logging/featureserviceschema.go). A known issue tracked in the community is that `get_online_features` against remote stores has dropped all but the first join key when multiple entity rows are passed ([issue #6488](https://github.com/feast-dev/feast/issues/6488)). This is a recurring pain point for users who model composite entities and rely on the remote online store adapter to forward the full request to a server-side feature server. ## Offline Stores The offline store abstraction lives in [sdk/python/feast/infra/offline_stores/offline_store.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/offline_stores/offline_store.py) and powers historical retrieval for training, backfills, and point-in-time joins. Recent releases have incrementally improved type fidelity and remote retrieval semantics: - v0.55.0 added `unix_timestamp_val` handling in `_serialize_val` so that timestamp columns round-trip cleanly through the offline store pipeline ([v0.55.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.55.0)). - v0.56.0 introduced date-wise historical data retrieval for remote offline stores ([v0.56.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.56.0)). - v0.57.0 and v0.60.0 improved Trino-to-Feast type mapping for `real`, `varchar`, `timestamp`, `decimal`, and `date` types ([v0.57.0](https://github.com/feast-dev/feast/releases/tag/v0.57.0), [v0.60.0](https://github.com/feast-dev/feast/releases/tag/v0.60.0)). - v0.59.0 added `get_table_query_string_with_alias()` for PostgreSQL subquery aliasing to keep generated SQL deterministic ([v0.59.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.59.0)). There is also an open feature request to add a zoned datetime feature type so that timezone information survives materialization and retrieval ([issue #6536](https://github.com/feast-dev/feast/issues/6536)). Until that lands, users typically store the timestamp and zone as two separate features. ## Materialization and Feature Retrieval Materialization is the bridge between the offline and online worlds. It reads feature values from the offline store for a time window and writes the latest value per entity into the online store. v0.58.0 added the ability to force fully qualified feature names for materialization operations, which removes ambiguity when multiple feature views share feature names ([v0.58.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.58.0)). v0.57.0 fixed the Materialize API for on-demand feature views so that transformations resolve correctly during the move from offline to online ([v0.57.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.57.0)). The typical retrieval flow is: ```mermaid flowchart LR A[Client] -->|entity_keys| B[Feast SDK / Feature Server] B -->|resolve join keys| C{Online Store} C -->|Redis| D[(Redis)] C -->|Remote| E[Remote Feast Server] C -->|Hybrid| F[Local Cache + Remote] C -->|Milvus| G[(Milvus Vector Index)] C -->|feature values| B B -->|rows per entity| A ``` The Go feature server in this repository logs the resolved schema for each retrieval request, including all join keys (deduplicated) and the inferred types, which makes debugging issues like the join-key dropping bug in [issue #6488](https://github.com/feast-dev/feast/issues/6488) much easier in production. Source: [go/internal/feast/server/logging/featureserviceschema.go:21-65](https://github.com/feast-dev/feast/blob/main/go/internal/feast/server/logging/featureserviceschema.go). ## Common Failure Modes - **Composite join keys over remote online store**: as in [issue #6488](https://github.com/feast-dev/feast/issues/6488), only the first join key may be forwarded by the remote adapter. Workaround is to use a single concatenated join key or to read directly from a colocated online store. - **Timezone loss on materialization**: timestamps lose zone information because zoned datetimes are not yet a first-class feature type ([issue #6536](https://github.com/feast-dev/feast/issues/6536)). Store zone-aware and timestamp columns separately until the feature lands. - **Type mapping drift on Trino and PostgreSQL backends**: addressed incrementally in v0.55–v0.60. Pin to a known-good minor version if you depend on a specific mapping behavior. ## See Also - [Project overview and architecture](https://github.com/feast-dev/feast/blob/main/README.md) - [Python feature server](https://docs.feast.dev/reference/feature-servers/python-feature-server) - [Feast Operator for Kubernetes deployment](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md) - [RAG and vector retrieval example](https://github.com/feast-dev/feast/blob/main/examples/rag-retriever/README.md) --- ## Feature Servers, Compute Engines, and Transformation Pipeline ### Related Pages Related topics: [Feast Overview, Architecture, and Core Concepts](#page-1), [Data Stores, Materialization, and Feature Retrieval](#page-2), [Deployment, Kubernetes Operator, Web UI, and Operations](#page-4)
Related Source Files The following source files were used to generate this page: - [sdk/python/feast/feature_server.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/feature_server.py) - [sdk/python/feast/offline_server.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/offline_server.py) - [sdk/python/feast/registry_server.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/registry_server.py) - [sdk/python/feast/transformation_server.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/transformation_server.py) - [sdk/python/feast/infra/mcp_servers/mcp_server.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/mcp_servers/mcp_server.py) - [sdk/python/feast/embedded_go/online_features_service.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/embedded_go/online_features_service.py) - [sdk/python/feast/infra/compute_engines/dag/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/infra/compute_engines/dag/README.md) - [infra/charts/feast/charts/feature-server/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/feature-server/README.md) - [infra/charts/feast/charts/transformation-service/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md) - [examples/mcp_feature_store/README.md](https://github.com/feast-dev/feast/blob/main/examples/mcp_feature_store/README.md) - [infra/feast-operator/README.md](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md)
# Feature Servers, Compute Engines, and Transformation Pipeline Feast exposes feature data to models and applications through a family of **servers** that share a common registry, online store, and offline store backplane. These servers, along with the compute engines that prepare data and the transformation pipeline that defines on-demand feature logic, form the runtime surface of the project. ## Overview of Feast Servers Feast ships four primary Python servers and one Go-embedded online service. Together they cover online retrieval, offline retrieval, registry access, on-demand transformation, and Model Context Protocol (MCP) integration. ### Python Feature Server The Python feature server is the default online retrieval service. It exposes `get_online_features` and related REST endpoints backed by a configurable online store. The Helm chart installs it as a separate pod whose image and ConfigMap can be overridden through `application-override.yaml` and `application-secret.yaml`. Source: [infra/charts/feast/charts/feature-server/README.md](). When run with the MCP extra, the feature server also registers FastAPI routes as MCP tools. The example `examples/mcp_feature_store/README.md` shows the required install command `pip install feast[mcp]` and the log line `INFO:feast.feature_server:MCP support has been enabled for the Feast feature server` that confirms the protocol layer is wired in. Source: [examples/mcp_feature_store/README.md](). ### Offline Feature Server and Registry Server The offline feature server (alpha) and registry server (alpha) expose historical retrieval and registry browsing over HTTP. These complement the Java feature server listed in the project roadmap and are useful for language-agnostic clients. ### Transformation Server The transformation server is a dedicated gRPC service that computes on-demand features. Its Helm chart is published with version `0.64.0` and uses the image `quay.io/feastdev/feature-transformation-server`. Source: [infra/charts/feast/charts/transformation-service/README.md](). It is invoked when feature views define UDFs that must run at request time rather than at materialization time. ### Embedded Go Online Service For low-latency deployments, Feast bundles a Go-based online features service that can be embedded directly inside a Python process. Source: [sdk/python/feast/embedded_go/online_features_service.py](). This path is commonly used when the host application already runs Python and wants to avoid an extra network hop. ## Server Topology and Request Flow The diagram below shows how a client request is routed through the available servers, the stores they touch, and where transformations are evaluated. ```mermaid flowchart LR Client[Client / Model] -->|REST/gRPC| FS[Python Feature Server] Client -->|MCP| MCP[MCP Server] FS --> Online[(Online Store)] FS --> TS[Transformation Server] FS --> Reg[(Registry)] MCP --> FS Client -->|Historical| OFS[Offline Feature Server] OFS --> Offline[(Offline Store)] Client -->|Browse| RS[Registry Server] RS --> Reg Embedded[Embedded Go Service] --> Online ``` Feature servers read feature view definitions from the registry, fetch latest values from the online store, and delegate any on-demand transformations to the transformation server. Source: [sdk/python/feast/feature_server.py](). Offline retrieval flows through the offline server, which queries the offline store and writes point-in-time-correct datasets. Source: [sdk/python/feast/offline_server.py](). ## Compute Engines and the DAG Abstraction Feast models feature engineering as a Directed Acyclic Graph (DAG). The DAG module lives under `sdk/python/feast/infra/compute_engines/dag/` and exposes five core abstractions: - `context.py` – `ExecutionContext` for managing per-feature-view execution state. - `node.py` – `DAGNode` representing an individual operation. - `value.py` – `DAGValue` for data passed between nodes. - `model.py` – data classes such as `DAGFormat`. - `plan.py` – `ExecutionPlan` that runs the graph. Source: [sdk/python/feast/infra/compute_engines/dag/README.md](). This abstraction lets the same logical plan run on different compute engines (local, Ray, Spark) without changing feature view definitions. ## Transformation Pipeline The transformation pipeline describes how raw inputs become features at retrieval time: 1. **Definition** – A `FeatureView` or `BatchFeatureView` declares one or more transformation fields. Recent releases added a `mode` field to the Transformation proto so that transformations are correctly serialized and persisted in the registry (v0.56.0). Source: [GitHub release v0.56.0](). 2. **Packaging** – The definition is bundled with the feature view and shipped to whichever server evaluates it. 3. **Evaluation** – At retrieval time the Python feature server forwards on-demand requests to the transformation server over gRPC. The transformation server returns computed values that are merged with the online store response. Source: [sdk/python/feast/transformation_server.py](). 4. **Static artifacts** – Templates such as `pytorch_nlp` demonstrate pre-loading lookup tables and small models at feature server startup, so each request avoids cold-start cost. Source: [sdk/python/feast/templates/pytorch_nlp/README.md](). ## Operational Notes The Feast Operator simplifies running these servers on Kubernetes by providing a `FeatureStore` CRD with persistence, registry topology, security, and batch job settings. The persistence guide (link in the operator README) covers file-based versus DB-backed online, offline, and registry stores. Source: [infra/feast-operator/README.md](). Community-reported issues worth noting: in v0.64.0 the `RedisOnlineStore` gained an `async_supported` property, and the operator CRD was updated with additional `feast init` templates. A separate issue ([#6488]()) describes `get_online_features` dropping all but the first join key on remote stores, which is relevant when validating online retrieval behavior. ## See Also - [README.md](https://github.com/feast-dev/feast/blob/main/README.md) – project overview and roadmap. - [infra/feast-operator/README.md](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md) – Kubernetes operator and FeatureStore CRD. - [java/serving/README.md](https://github.com/feast-dev/feast/blob/main/java/serving/README.md) – Java feature server developer guide. - [sdk/python/feast/templates/ray/README.md](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/templates/ray/README.md) – Ray compute engine template. --- ## Deployment, Kubernetes Operator, Web UI, and Operations ### Related Pages Related topics: [Feast Overview, Architecture, and Core Concepts](#page-1), [Feature Servers, Compute Engines, and Transformation Pipeline](#page-3)
Related Source Files The following source files were used to generate this page: - [infra/feast-operator/README.md](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md) - [infra/feast-operator/config/crd/bases/feast.dev_featurestores.yaml](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/config/crd/bases/feast.dev_featurestores.yaml) - [infra/charts/feast/values.yaml](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/values.yaml) - [infra/charts/feast-feature-server/values.yaml](https://github.com/feast-dev/feast/blob/main/infra/charts/feast-feature-server/values.yaml) - [infra/charts/feast/charts/feature-server/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/feature-server/README.md) - [infra/charts/feast/charts/transformation-service/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md) - [ui/README.md](https://github.com/feast-dev/feast/blob/main/ui/README.md) - [java/serving/README.md](https://github.com/feast-dev/feast/blob/main/java/serving/README.md) - [examples/operator-quickstart/README.md](https://github.com/feast-dev/feast/blob/main/examples/operator-quickstart/README.md) - [examples/rbac-remote/README.md](https://github.com/feast-dev/feast/blob/main/examples/rbac-remote/README.md) - [sdk/python/feast/cli/cli.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/cli/cli.py) - [README.md](https://github.com/feast-dev/feast/blob/main/README.md)
# Deployment, Kubernetes Operator, Web UI, and Operations ## Overview Feast ships with several first-class mechanisms for production deployment and day-2 operations: a Kubernetes **Operator** that reconciles `FeatureStore` custom resources, a composable set of **Helm charts** for the core components, a Python **Web UI** for browsing feature repositories, and a **CLI** that drives apply/materialize/serve workflows. Source: [README.md](https://github.com/feast-dev/feast/blob/main/README.md). The recommended path to production is to install the Operator once per cluster and then declare feature stores declaratively as Kubernetes resources. The Operator currently uses the `0.64.0` chart version, as reflected in `appVersion: v0.64.0` badges on the published sub-charts (Source: [infra/charts/feast/charts/feature-server/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/feature-server/README.md), [infra/charts/feast/charts/transformation-service/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md)). ## Feast Kubernetes Operator The Operator lives under `infra/feast-operator/` and exposes a single CRD, `FeatureStore`, defined in `feast.dev_featurestores.yaml`. A `FeatureStore` spec can declare `feastProjectDir` (git clone or `feast init` template), persistence choices (file path, PVC, or DB for offline/online/registry), serving settings (workers, log level, Prometheus metrics, MCP), registry topology (local, remote, cross-namespace via `feastRef`), security (RBAC roles vs OIDC), and batch/jobs settings (`batchEngine` ConfigMap, `cronJob` for materialization) (Source: [infra/feast-operator/README.md](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/README.md)). The full CRD surface is documented in the project's `docs/api/markdown/ref.md`, and the canonical types are defined in Go under `infra/feast-operator/api/v1/featurestore_types.go` (Source: [infra/feast-operator/config/crd/bases/feast.dev_featurestores.yaml](https://github.com/feast-dev/feast/blob/main/infra/feast-operator/config/crd/bases/feast.dev_featurestores.yaml)). A guided quickstart is provided in `examples/operator-quickstart/` with notebooks for install, demo, and uninstall steps (Source: [examples/operator-quickstart/README.md](https://github.com/feast-dev/feast/blob/main/examples/operator-quickstart/README.md)). The operator bundles the missing `feast init` templates and persistence documentation, addressing a long-standing operator-to-CLI parity gap (cf. the v0.64.0 release notes: "Add missing feast init templates to operator CRD and enhance persistence documentation") (Source: [v0.64.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.64.0)). ```mermaid flowchart LR User[User / Platform Team] -->|kubectl apply| CR[FeatureStore CR] CR --> Op[Feast Operator Controller] Op -->|reconciles| FS[FeatureStore Pods
online + offline + registry] Op -->|provisions| PVC[(PVC / DB / Object Store)] FS -->|gRPC / REST| Client[Feast Python/Go/Java SDK] Client -->|get_online_features| Online[(Online Store
Redis, etc.)] Client -->|get_historical_features| Offline[(Offline Store
BigQuery, Snowflake, ...)] ``` ## Helm Charts and Image Topology The umbrella chart `infra/charts/feast/` composes three sub-charts — `feature-server` (the Java online serving service), `transformation-service` (the gRPC on-demand transformation server), and supporting persistence objects — each pinned to `0.64.0`. The `feature-server` chart exposes `application-override.yaml` and `application-secret.yaml` precedence rules, allowing operators to layer overrides via ConfigMap or Secret (Source: [infra/charts/feast/charts/feature-server/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/feature-server/README.md)). The transformation chart exposes `envOverrides`, `podLabels`, `replicaCount`, `nodeSelector`, and `service.grpc.port/nodePort` for fan-out tuning (Source: [infra/charts/feast/charts/transformation-service/README.md](https://github.com/feast-dev/feast/blob/main/infra/charts/feast/charts/transformation-service/README.md)). For developers who prefer raw Helm without the Operator, the standalone `infra/charts/feast-feature-server/` chart is maintained separately (Source: [infra/charts/feast-feature-server/values.yaml](https://github.com/feast-dev/feast/blob/main/infra/charts/feast-feature-server/values.yaml)). A working Redis + MinIO end-to-end walkthrough is available in `examples/python-helm-demo/`, including `kubectl port-forward` recipes and the `REDIS_PASSWORD` extraction pattern from the Bitnami Redis secret. ## Feast Web UI The experimental Web UI (`ui/`) is a Create React App that can be launched in three ways: through the `feast ui` CLI to inspect the local feature repository, by importing `@feast-dev/feast-ui` as a module into a host React application, or by running the full standalone build (Source: [ui/README.md](https://github.com/feast-dev/feast/blob/main/ui/README.md)). When imported as a module, the host app supplies a `projects-list.json` describing the repositories to render; optional peer dependencies are `@elastic/eui` and `@emotion/react` for theming. Community issues track regressions across upgrades — e.g. the UI rendered incorrectly after a 0.63.0 → 0.64.0 demo upgrade (issue [#6539](https://github.com/feast-dev/feast/issues/6539)). Operators rolling forward should pin the UI version and re-validate the demo flow against the published `0.64.0` charts. ## Operations: CLI, Serving, and RBAC The Python CLI in `sdk/python/feast/cli/cli.py` is the operational entrypoint: it implements `feast init`, `feast apply`, `feast materialize`, `feast materialize-incremental`, `feast serve`, and `feast ui`. The `feast serve` command launches the Python feature server, while the Java serving module under `java/serving/` provides the gRPC/HTTP online serving implementation packaged as an executable jar via Maven (`mvn -f java/serving/pom.xml package`) (Source: [java/serving/README.md](https://github.com/feast-dev/feast/blob/main/java/serving/README.md), [sdk/python/feast/cli/cli.py](https://github.com/feast-dev/feast/blob/main/sdk/python/feast/cli/cli.py)). Recent v0.64.0 work added an `async_supported` property to `RedisOnlineStore` to unblock async materialization paths (Source: [v0.64.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.64.0)). For multi-tenant deployments, RBAC authorization can be enforced at the registry boundary. The `examples/rbac-remote/` example ships three personas — `feast-admin-sa`, `feast-user-sa`, and `feast-unauthorized-sa` — with corresponding roles (`feast_admin_permission`, `feast_user_permission`, `feast_unauthorized_permission`) and an `auth_type: kubernetes` block in `feature_store.yaml` (Source: [examples/rbac-remote/README.md](https://github.com/feast-dev/feast/blob/main/examples/rbac-remote/README.md)). Common production failure modes discussed in the community include dropped join keys on remote online stores ([issue #6488](https://github.com/feast-dev/feast/issues/6488)) and missing readiness checks on the REST registry, both of which are addressed by targeted fixes in recent releases (Source: [v0.61.0 release notes](https://github.com/feast-dev/feast/releases/tag/v0.61.0)). ## See Also - [Feast main README](https://github.com/feast-dev/feast/blob/main/README.md) - [Feast Operator configuration guides](https://docs.feast.dev/how-to-guides/feast-operator) - [Java Serving developer guide](https://github.com/feast-dev/feast/blob/main/java/serving/README.md) - [Operator quickstart examples](https://github.com/feast-dev/feast/tree/main/examples/operator-quickstart) - [RBAC remote example](https://github.com/feast-dev/feast/tree/main/examples/rbac-remote) --- --- ## Pitfall Log Project: feast-dev/feast Summary: Found 8 structured pitfall item(s), including 1 high/blocking item(s). Top priority: Installation risk - Installation risk requires verification. ## 1. Installation risk - Installation risk requires verification - Severity: high - Evidence strength: source_linked - Finding: Project evidence flags a installation risk. Review the linked source before relying on this workflow. - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: community_evidence:github | https://github.com/feast-dev/feast/issues/6488 ## 2. Capability evidence risk - Capability evidence risk requires verification - Severity: medium - Evidence strength: source_linked - Finding: README/documentation is current enough for a first validation pass. - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: capability.assumptions | https://github.com/feast-dev/feast ## 3. Maintenance risk - Maintenance risk requires verification - Severity: medium - Evidence strength: source_linked - Finding: Project evidence flags a maintenance risk. Review the linked source before relying on this workflow. - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: community_evidence:github | https://github.com/feast-dev/feast/issues/6539 ## 4. Maintenance risk - Maintenance risk requires verification - Severity: medium - Evidence strength: source_linked - Finding: Project evidence flags a maintenance risk. Review the linked source before relying on this workflow. - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: evidence.maintainer_signals | https://github.com/feast-dev/feast ## 5. Security or permission risk - Security or permission risk requires verification - Severity: medium - Evidence strength: source_linked - Finding: no_demo - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: downstream_validation.risk_items | https://github.com/feast-dev/feast ## 6. Security or permission risk - Security or permission risk requires verification - Severity: medium - Evidence strength: source_linked - Finding: no_demo - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: risks.scoring_risks | https://github.com/feast-dev/feast ## 7. Maintenance risk - Maintenance risk requires verification - Severity: low - Evidence strength: source_linked - Finding: issue_or_pr_quality=unknown。 - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: evidence.maintainer_signals | https://github.com/feast-dev/feast ## 8. Maintenance risk - Maintenance risk requires verification - Severity: low - Evidence strength: source_linked - Finding: release_recency=unknown。 - User impact: May increase setup, validation, or first-run risk for the user. - Evidence: evidence.maintainer_signals | https://github.com/feast-dev/feast