Multi-Cluster Kubeflow

This is project is proof-of-concept for how one could go about running Kubeflow notebooks on multiple clusters.

By using Admiralty for multi-cluster scheduling, we can achieve a cloud bursting topology where a primary "Data Center" cluster can be used to run the istio Control Plane, the Kubeflow Components as well provide a fixed compute resource amount, while a secondary "Cloud" cluster can meshed with it via Istio provide elastic compute capacity to the platform.

In this repo, 2 local kind clusters are created to demonstrate the above scenario:

.

.

---

Demo

demo.mp4

Setup

Prerequisites

Ensure all dependencies are available in your systemd:

make dependencies

Install

Single command

make all

Step by step equivalent

make up # create clusters and configure admiralty
make istio # configure istio mesh
make kubeflow # install, configure and expose kubeflow

Testing

make test # test istio, admiralty and kubeflow setups

My configuration

• OS: Ubuntu 22.04, Arch: amd64
• make: 4.3
• bash: 5.1.16
• kind: kind v0.23.0 go1.21.10 linux/amd64
• istio/istioctl: 1.23.2

---

Sources

• https://admiralty.io/docs/
• https://istio.io/v1.23/docs/setup/install/multicluster/primary-remote/
• https://github.com/mlkmhd/istio-multi-cluster-initializer/tree/8aad469ff6f5ece7e01f9c1f9e76375ac1c5ece9
• https://github.com/kubeflow/manifests/tree/v1.9.0