Simplify Kubeflow Deployment with kubeflow-chart: A Step‑by‑Step Guide

Background

Kubeflow is the de‑facto MLOps platform on Kubernetes, but vanilla Kubeflow requires extensive manual configuration, many dependent components, and considerable operational effort, especially in on‑premise clusters that already run Istio, Knative, or Dex.

Key Challenges of Vanilla Kubeflow

Complex, multi‑step deployment using Kustomize overlays.

Numerous tightly coupled components increase maintenance overhead.

Poor usability of core features makes application development and migration difficult.

No native Chinese language support and limited documentation for local mirrors.

kubeflow‑chart Overview

The open‑source kubeflow‑chart (maintained by Alauda) packages Kubeflow together with extensions such as MLflow, SQLFlow, kfpdist, Elyra, Volcano and gang‑scheduler. All common configuration items—image registry, authentication, default service accounts, and component enable/disable switches—are centralized in a single values.yaml. A China‑specific values‑cn.yaml provides alternative image sources to avoid pull‑rate limits.

Deployment Procedure (Minikube Example)

helm repo add alauda https://alauda.github.io/kubeflow-chart

# Install prerequisite components

helm install istio alauda/istio

helm install cert-manager alauda/certmanager

helm install my-kubeflow alauda/kubeflow

For users in mainland China, replace the default charts with the China‑specific value files:

wget -O values-istio-cn.yaml https://raw.githubusercontent.com/alauda/kubeflow-chart/charts/istio/values-cn.yaml
helm install istio alauda/istio -f values-istio-cn.yaml

wget -O values-certm-cn.yaml https://raw.githubusercontent.com/alauda/kubeflow-chart/charts/certmanager/values-cn.yaml
helm install cert-manager alauda/certmanager -f values-certm-cn.yaml

wget https://raw.githubusercontent.com/alauda/kubeflow-chart/main/values-cn.yaml
helm install my-kubeflow alauda/kubeflow -f values-cn.yaml

After the installation, forward the Istio ingress gateway and open the UI at http://localhost:8080/ using the default credentials [email protected] / 12341234 :

kubectl port-forward svc/istio-ingressgateway -n istio-system --address=0.0.0.0 8080:80

Integrated Extensions

SQLFlow – Enables model training and inference directly from SQL statements inside Jupyter notebooks. Supports MySQL, TiDB, Hive and other relational databases.

kfpdist – A KFP SDK plugin that adds a ParallelFor construct, allowing a single workflow step to launch a distributed training job without creating separate TFJob resources.

Elyra – Provides a visual authoring layer on top of Jupyter for building data‑science pipelines.

Volcano & gang‑scheduler – Advanced AI‑task schedulers offering resource‑aware placement, priority queues, and pre‑emptive scheduling.

MLflow – Adds richer experiment tracking, automatic logging of parameters, metrics, and model artifacts.

Roadmap and Future Enhancements

Multi‑tenant isolation and vGPU support.

High‑availability deployment across multiple regions.

Chinese‑localized UI and documentation.

Comprehensive tutorials, monitoring, and alerting integrations.

Conclusion

kubeflow‑chart delivers a Helm‑based, one‑click installation path for a production‑grade MLOps stack on any Kubernetes cluster. By bundling Kubeflow with SQLFlow, kfpdist, Elyra, Volcano, and MLflow, it mitigates the usability and operational gaps of the vanilla Kubeflow distribution.

Repository: https://github.com/alauda/kubeflow-chart