Skip to main content

Online Feature Store - Release Notes

Version 1.0.0 🚀​

Release Date: June 2025
Status: General Availability (GA)

We're excited to announce the first stable release of the BharatML Online Feature Store - a high-performance, production-ready feature serving system designed for machine learning workloads.

🎯 What's New​

Core Feature Store Engine​

  • Ultra-Low Latency: Achieve sub-10ms P99 response times for real-time inference
  • High Throughput: Tested and validated at 1M+ requests per second with 100 IDs per request
  • Multi-Entity Support: Serve features for multiple entity types (users, transactions, products, etc.)
  • Batch Retrieval: Efficient bulk feature fetching for real-time inference and incremental/online training workloads

Advanced Data Type Support​

Complete support for all ML-relevant data types:

Data TypeVariantsOptimizations
Integersint8, int16, int32, int64Varint encoding, bit packing
Floatsfloat8, float16, float32, float64IEEE 754 compliant storage
StringsVariable lengthPascal string encoding
BooleansBit-packed8x memory compression
VectorsAll above typesContiguous memory layout

Multi-Database Architecture​

Flexible backend storage with optimized drivers:

  • 🔥 Scylla DB: Ultra-high performance NoSQL (recommended for production)
  • ⚡ Dragonfly: Modern Redis alternative with better memory efficiency
  • 📊 Redis: Standard in-memory store for development environments

🚀 Performance & Optimization​

PSDB v2 Serialization Format without compression​

Our proprietary Permanent Storage Data Block format delivers:

  • 35% faster serialization than Protocol Buffers
  • 100.0-102.2% size efficiency (near raw data size)
  • 93% fewer allocations than Apache Arrow (4 vs 66 allocs/op)
  • 975 MB/s throughput capacity

Memory Management​

  • Object Pooling: Zero-allocation feature retrieval with PSDBPool
  • Connection Pooling: Optimized database connection reuse
  • Buffer Management: Pre-allocated buffers for serialization operations
  • Smart Caching: Configurable TTL-based feature caching

Compression Support​

Intelligent compression with multiple algorithms:

  • ZSTD: Maximum compression for bandwidth-constrained environments
  • Auto-Fallback: Intelligent selection based on data characteristics

🛠️ APIs & SDKs​

gRPC API​

High-performance, language-agnostic interface:

service FeatureStoreService {
    rpc RetrieveFeatures(Query) returns (QueryResult);
    rpc RetrieveDecodedFeatures(Query) returns (DecodedQueryResult);
    rpc PersistFeatures(PersistFeaturesRequest) returns (Result);
}

Go SDK v1.0.0​

Native Go client with enterprise features:

  • Type-Safe API: Strongly typed interfaces and data structures
  • Connection Management: Configurable timeouts, TLS, and pooling
  • Batch Processing: Configurable batch sizes for bulk operations
  • Metrics Integration: Built-in timing and count metrics
  • Authentication: Caller ID and token-based security

Python SDK Collection v1.0.0​

Three specialized Python packages for different ML workflows:

bharatml_commons - Common utilities and protobuf definitions:

  • HTTP Client: Feature metadata operations
  • Protobuf Support: Generated Python definitions for all APIs
  • Utility Functions: Column cleaning and feature processing

spark_feature_push_client - Apache Spark-based data pipeline:

  • Batch ETL: Large-scale data processing with Spark
  • Kafka Integration: Protobuf serialization and Kafka publishing
  • Multi-Source Support: Hive, Delta, Parquet, Cloud Storage

grpc_feature_client - High-performance gRPC client:

  • Real-time Operations: Direct persist/retrieve API access
  • Low Latency: Optimized for model inference workflows
  • Type Safety: Strongly typed Python interfaces

RESTful Interface​

HTTP API for web applications:

  • Health Endpoints: Built-in monitoring and status checks

🔧 Enterprise Features​

Production Readiness​

  • Health Checks: /health/self endpoints for probing
  • Graceful Shutdown: Clean resource cleanup with configurable timeouts
  • Structured Logging: Formatted logs with configurable levels
  • Signal Handling: SIGTERM/SIGINT support for container environments

Monitoring & Observability​

  • DataDog Integration: Built-in metrics collection and reporting
  • Prometheus Compatibility: Standard metrics format support
  • Custom Metrics: Request rates, latencies, error rates, and business metrics
  • Distributed Tracing [untested]: Request flow visibility across services

Data Management​

  • TTL Support: Automatic feature expiration
  • Feature Versioning: Schema evolution with backward compatibility
  • Bulk Operations: Efficient batch read/write with configurable sizes

🏗️ Deployment & Configuration​

Container Support​

  • Docker Images: Multi-architecture support (amd64, arm64)

🔄 Compatibility​

Supported Go Versions​

  • Minimum: Go 1.22.0
  • Recommended: Go 1.22.8+

Database Compatibility​

DatabaseVersionStatusNotes
Scylla DB5.0+✅ RecommendedOptimal performance
Dragonfly1.0+✅ SupportedMemory efficient
Redis6.0+✅ DevelopmentLimited scale

🐛 Known Issues​

Current Limitations​

  1. Large Vector Support: Vectors >10MB may experience increased latency

Workarounds​

  1. Vector Chunking: Split large vectors into smaller segments

💾 Download & Installation​

Container Images​

# Pull the latest image
docker pull ghcr.io/meesho/onfs-api-server:latest
docker pull ghcr.io/meesho/onfs-consumer:latest
docker pull ghcr.io/meesho/horizon:latest
docker pull ghcr.io/meesho/trufflebox-ui:latest

Arch Supported​

  • Linux (amd64)
  • Linux (arm64)
  • macOS (Intel)
  • macOS (Apple Silicon)

Checkout Packages

Source Code​

git clone https://github.com/Meesho/BharatMLStack.git
cd BharatMLStack/online-feature-store
git checkout release/1.0.0

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!