System architecture

Charmed Feast provides a feature store solution tailored for MLOps workflows on Charmed Kubeflow. It leverages Juju to manage Feast components and integrates seamlessly with Kubeflow Notebooks for end-to-end machine learning use cases.

Feast Architecture Diagram

Charmed Feast – Feature store architecture

This architecture deploys Feast with PostgreSQL-based data stores, and integrates with Kubeflow to provide a consistent developer experience for managing features during training and inference.

Key features

  • PostgreSQL used for offline, online, and registry stores.

  • Feature store configuration managed as a Kubernetes (K8s) secret.

  • Seamless Notebook integration via PodDefault.

  • Optional User Interface (UI) access through Kubeflow Dashboard and Istio Ingress.

Components

The following components constitute a complete Charmed Feast deployment:

Feast Integrator charm

It acts as the central orchestrator by:

  • Establishing relations with PostgreSQL charms, including offline, online, and registry.

  • Rendering and managing the feature_store.yaml configuration file.

  • Creating a K8s secret and PodDefault to share this configuration with Notebooks.

  • Relating with the Resource Dispatcher to propagate configurations to the user namespace.

See Feast Integrator for more details about this charm.

Feast UI charm

It provides a web interface for browsing Feast objects, such as feature views and entities, by:

  • Retrieving feature_store.yaml from Feast Integrator.

  • Running the Feast UI as a Pebble service.

  • Sending a DashboardLink to the Kubeflow dashboard.

  • Using Istio Ingress for external access.

See Feast UI for more details about this charm.

PostgreSQL charms

Three PostgreSQL deployments are used:

  • Offline store: Stores historical feature data for training.

  • Online store: Serves features at low latency for inference.

  • Registry: Stores metadata about feature definitions and entities.

All PostgreSQL charms communicate with Feast via the postgresql_client interface. See PostgreSQL for more details about this charm.

Resource Dispatcher charm

This charm propagates feature_store.yaml and PodDefault to the user namespace, ensuring Notebooks can access Feast configurations.

User Notebooks

Kubeflow Notebooks are the main interface for users to interact with Feast. As a user, you can:

  • Run feast apply to register features.

  • Run feast materialize to load data into the online store.

  • Retrieve historical and online features for training or inference.

Charmed Kubeflow integration

Charmed Feast is designed to tightly integrate with Charmed Kubeflow:

  • K8s secrets and PodDefaults: The Feast Integrator charm creates a K8s secret containing feature_store.yaml, and a PodDefault to mount it into user Notebooks.

  • UI access: The Feast UI charm integrates with: - The Kubeflow Dashboard charm to add Feast UI to the sidebar. - The Istio Pilot charm to enable Ingress routing for external access.

  • Notebook support: you can install the Feast SDK as follows:

    pip install feast feast[postgres]

Once done, you can interact with Feast directly using your Notebook terminal.

Note

You must materialize features before retrieving them for online inference. See Feast quickstart guide for more information.