Skip to main content

BharatMLStack - Online Feature Store (OnFS)

The Online Feature Store (OnFS) is part of BharatMLStack, designed to support real-time ML workloads through low-latency feature retrieval and flexible feature ingestion pipelines. It ensures that features generated offline or online are immediately accessible for inference.

BharatMLStack's Online-feature-store Architecture

🧩 Key Components​

1. Data Ingestion Paths​

a. Direct Push from Feature Engineering Jobs​

  • Source: Apache Spark
  • Client: spark_feature_push_client
  • Flow: Features are pushed directly to Kafka.

b. Push from Offline Feature Store​

  • Source: Delta Lake, GCS, or S3
  • Flow: Scheduled notebooks (push_features_to_online-feature-stores.ipynb) push to Kafka using the same spark_feature_push_client.
  • Source: Flink streaming jobs
  • Client: custom-producer
  • Flow: Real-time features sent to Kafka.

2. Message Queue: Kafka​

Kafka serves as a decoupled buffer between producers (push clients) and consumers (OnFS ingestion workers), ensuring durability and backpressure handling.

3. Core Components​

🧠 Horizon Control Plane​

  • Manages config distribution and metadata orchestration.
  • Stores schemas, feature group mappings, job configurations in etcd.

🔍 Trufflebox UI​

Frontend interface for managing the ML Feature Store ecosystem:

  • Feature Catalog – Browse, search, and inspect registered features and groups.
  • Store and Job Registry – View and manage ingestion jobs, feature store states, and lineage.
  • Admin Ops – Approve or reject feature group pushes and schema edits.
  • Designed for use by Data Scientists, MLEs, and Platform Admins.

⚙️ OnFS-Consumers​

  • Kafka consumers that read and validate feature messages.
  • Responsible for persisting features to online databases (Redis, ScyllaDB, DragonflyDB).

🚀 OnFS API Server​

  • gRPC server exposing interfaces for:
    • Real-time feature persistence.
    • Low-latency feature retrieval.
  • Access controlled and schema-validated via etcd.

4. Online Databases​

Stores real-time features for high-performance retrieval:

  • DragonflyDB
  • Redis
  • ScyllaDB

5. Clients for Serving​

Applications use client SDKs to fetch features:

  • Go SDK: go-sdk
  • Python SDK: grpc-feature-client
  • Used in backend inference apps to request features using entity keys.

6. Observability​

  • Prometheus – Metrics collection (e.g., ingest lag, QPS, latency).
  • Grafana – Dashboard for platform health, feature access, ingestion success/failure.

💻 Supported Environments​

  • Kubernetes (K8s)
  • Google Kubernetes Engine (GKE)
  • Amazon EKS

👥 Target Users​

UserRole
Data ScientistsBrowse features, define jobs, approve/reject changes via Trufflebox UI
MLEsDevelop and push features using Spark/Flink/notebooks
Infra AdminsManage store lifecycle, metadata, and approvals
Backend DevsUse SDKs to retrieve features in Go/Python inference services

✅ Benefits​

  • Unified real-time and offline ingestion.
  • Low-latency inference-ready features.
  • Config-driven orchestration.
  • Built-in approval workflows via Trufflebox.
  • Scalable across thousands of entities and feature groups.

Contributing​

We welcome contributions from the community! Please see our Contributing Guide for details on how to get started.

Community & Support​

License​

BharatMLStack is open-source software licensed under the BharatMLStack Business Source License 1.1.

Built with ❤️ for the ML community from Meesho
If you find this useful, ⭐️ the repo — your support means the world to us!