Meesho open-sources its homegrown ML stack for 'Bharat scale'

What it does

BharatMLStack is a full-suite ML infrastructure platform that bundles a sub-10ms online feature store, a DAG-based real-time inference orchestrator called Inferflow, a vector search engine (Skye), a Rust-powered math kernel (Numerix), and a control plane (Horizon) that ties them together through a web UI. It targets teams running real-time personalization, fraud detection, or image search at high throughput, and claims 60–70% lower infrastructure costs compared to hyperscaler managed services. Everything is Kubernetes-native and, according to the README, runs across public cloud, on-prem, and edge.

The interesting bit

The stack is not a paper tiger: Meesho says it already drives 2.4M QPS through its feature store and 1M+ QPS for model inference in production. The most unusual ingredient is the explicit focus on “Bharat scale” economics—optimizing for cost and latency in markets where infrastructure spend is scrutinized far more than in Silicon Valley.

Key highlights

• Online feature store serving 2.4M QPS (batch of 100 ID lookups) with sub-10ms latency.
• Inferflow for DAG-based real-time inference orchestration and Skye for vector similarity search.
• Numerix, a Rust-powered compute engine for high-performance matrix operations.
• Kubernetes-native control plane (Horizon) and web console (TruffleBox UI) for feature registry and approvals.
• Go and Python SDKs to hook the feature store and inference logging into existing code.

Caveats

• The project is released under the Business Source License 1.1, so it is source-available but not fully open source in the OSI sense; future commercial use may face restrictions.
• Performance and cost claims (e.g., 60–70% savings, 99.99% uptime) are self-reported from Meesho’s internal deployment; independent benchmarks are not provided.
• Several components, such as the Interaction Store, currently lack version tags and published documentation links.

Verdict

Teams running high-QPS, low-latency ML in production—especially those trying to escape hyperscaler bills—should evaluate this. If you are looking for a fully OSI-licensed stack or a managed SaaS, look elsewhere.