> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# Architecture and Security

## Architecture

As a Snowflake native app, Kumo and its various required binaries, container images, and packages are available as a single package in the Snowflake Marketplace. Kumo runs multiple containers (for example, the UI, REST, workers, and trainer) that perform various functions to enable users to connect, process data, and build ML models for predictive analytics.

Kumo’s architecture consists of two key layers:

**Control Plane**

The control plane is a collection of services that includes the metadata manager, a compiler that both translates predictive queries to an execution plan for the graph model and is used to generate predictions, and workflow orchestrators to coordinate various activities across Kumo.

**Predictive Query Engine**

The **Predictive Query Engine** is a heterogeneous distributed system that processes predictive queries. It contains multiple components, each responsible for a specific function in the ML pipeline.

The **Data Engine** processes the input relational data using [Snowpark APIs](https://docs.snowflake.com/en/developer-guide/snowpark/index) and generates graph and training data used to build the GNN model. These outputs are materialized to an intermediate Snowflake stage, owned by the client.

The **Graph Engine** loads graphs into the Graph Store and node attributes into the Column Store. Its primary role is to serve subgraph requests and node attributes for GNN training and batch inference.

* The Graph Store contains all entities in the data warehouse (rows).
* The Column Store contains all the attributes about the entities (columns).

The **GPU trainer** is responsible for GNN training and batch inference. It uses the available GPU compute in Snowflake to build the core graph model.

Kumo’s **GNN models** learn from graph-structured data and use the leading open-source framework, PyTorch Geometric, for **model execution**. Kumo uses a variety of graph neural network architectures and training procedures especially designed for learning on relational databases.

## Security

The Snowflake native app supports Kumo's end-to-end machine learning (ML) platform. Kumo enables enterprises to use predictive analytics to solve prediction problems: data scientists register data sources and issue SQL-like predictive queries that specify ML tasks. Kumo then automates feature preparation, label engineering, training dataset creation, model optimization, and MLOps.

The Kumo native app for Snowflake provides security benefits for your organization, because all data resides within your Snowflake environment.

![Security architecture diagram of Kumo running as a Snowflake Native App on SPCS](https://fdr-prod-docs-files-public.s3.us-east-1.amazonaws.com/nvidia-sdgm.docs.buildwithfern.com/e80b2f7cfb87726b5f776ae7eb4efd3c6b62e6c6f9379de7c3f81461ea2ff814/img/spcs_native_app/spcs_security.png?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIA6KXJSKKNFOCF7G4B%2F20260920%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20260920T154543Z&X-Amz-Expires=604800&X-Amz-Signature=4646b0c196178288700e322101c72cee6ceb66b246a9e594ef4ec7a1b4f92935&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject)