Pulumi EKS ML Infrastructure

An opinionated library for multi-tenant, multi-region Machine Learning platforms on AWS.

This repository provides a modular set of Pulumi components (pulumi_eks_ml) to spin up multi-tenant, multi-region ML infrastructure with minimal pain.

💡 Philosophy

This project treats infrastructure as a composable library. Instead of one giant deployment, you get modular building blocks (VPC, EKS, GPU Node Pools) that you can assemble into your own topology.

Whether it's a single cluster for testing or a global mesh for distributed workloads, you can define your architecture once in Python, then deploy identical copies across different environments thanks to Pulumi stacks.

Architectural examples with pulumi_eks_ml

⚡ Quickstart

Use the starter project as the fastest path to a working EKS cluster.

# __main__.py
import pulumi

from pulumi_eks_ml import eks, eks_addons, vpc

main_region = pulumi.Config("aws").require("region")
cfg = pulumi.Config()
deployment_name = f"{pulumi.get_project()}-{pulumi.get_stack()}"
node_pools_config = cfg.require_object("node_pools")

node_pools = [eks.NodePoolConfig.from_dict(pool) for pool in node_pools_config]

vpc_resource = vpc.VPC(
    name=f"{deployment_name}-vpc",
    cidr_block="10.0.0.0/16",
    setup_internet_egress=True,
)

cluster = eks.EKSCluster(
    f"{deployment_name}-cls",
    vpc_id=vpc_resource.vpc_id,
    subnet_ids=vpc_resource.private_subnet_ids,
    node_pools=node_pools,
)

eks.cluster.EKSClusterAddonInstaller(
    f"{deployment_name}-addons",
    cluster=cluster,
    addon_types=eks_addons.recommended_addons(),
)

pulumi.export("vpc_id", vpc_resource.vpc_id)
pulumi.export("cluster_name", cluster.cluster_name)

uv sync --dev
cd projects/starter
pulumi stack init dev
pulumi config set aws:region us-west-2
uv run pulumi up

🚀 Key Features

• ML-Optimized Compute: Pre-configured EKS clusters with Karpenter for autoscaling (Spot/On-Demand) and NVIDIA GPU drivers ready to go.
• Global Networking: Easy Multi-Region connectivity with Hub-and-Spoke or Full Mesh VPC peering topologies.
• Opinionated Add-ons for ML: Built-in support for ALB Controller, EBS/EFS CSI drivers, FluentBit, Metrics Server, etc...
• Secure network with Tailscale: Secure network with Tailscale for VPN access, in additional to public/private subnet isolation.
• SkyPilot Multi-Tenant Platform: Opinionated deployment of SkyPilot for multi-tenant, multi-region AI workloads.

📂 Repository Structure

• pulumi_eks_ml/: The core Python library containing reusable infrastructure components.
• projects/: Reference implementations and live infrastructure code.
  • starter/: A simple single-region EKS cluster.
  • multi-region/: A full-mesh global network connecting clusters across regions.
  • skypilot-multi-tenant/: A SkyPilot platform with isolated data planes for multiple teams.

🛠 Getting Started

Prerequisites

• Pulumi CLI
• Python 3.12+
• uv (recommended)

1. Install & Setup

# Clone the repo
git clone https://github.com/Roulbac/pulumi-eks-ml.git
cd pulumi-eks-ml

# Install dependencies
uv sync --dev

2. Deploy a Project

Navigate to one of the reference projects to see it in action.

cd projects/starter

# Initialize your stack (e.g., dev)
pulumi stack init dev

# Deploy
uv run pulumi up

For custom infrastructure, create a new folder in projects/, import pulumi_eks_ml, and define your topology (see projects/starter/__main__.py for a template).

🧪 Testing

We include both unit and integration tests (using LocalStack).

# Run Unit Tests
uv run pytest -vv tests/unit

# Run Integration Tests
uv run pytest -vv tests/integration

📄 License

MIT