Mastering Kubeflow: Deploy AI Workflows on Kubernetes Step‑by‑Step

What is Kubeflow?

Kubeflow is a Google‑originated machine‑learning platform built on Kubernetes, enabling definition of resources such as TFJob and allowing distributed model training similar to deploying regular applications.

Background: ML Model Service Lifecycle

Before Kubeflow, a typical ML model deployment passes through data cleaning, dataset splitting, training, model validation, large‑scale training, model export, service launch, and monitoring. Frameworks like TensorFlow address core training but leave many production‑level concerns such as data collection, feature extraction, resource management, model serving, configuration, storage, and logging.

Kubeflow Core Components

Jupyter multi‑tenant Notebook service

TensorFlow / PyTorch engines

Seldon for model deployment on Kubernetes

TF‑Serving for online TensorFlow model serving with versioning

Argo workflow engine

Ambassador API gateway

Istio for micro‑service management and telemetry

Ksonnet for deploying Kubernetes resources

Why Use Kubeflow on Kubernetes?

Native resource isolation

Cluster‑wide automated management

Automatic CPU/GPU scheduling

Support for various distributed storage backends

Mature monitoring and alerting

These features let you compose the ML pipeline as micro‑services and deploy them containerized for high availability and easy scaling.

Installation Guide

Server Requirements

GPU: Nvidia‑Tesla‑K80

Network: 10 GbE (Gigabit may become a bottleneck for large datasets)

CephFS Service

Use 10 GbE and co‑locate Ceph with the Kubernetes cluster to avoid high latency.

Software Versions

Kubernetes v1.12.2 (requires kube‑dns)

Kubeflow v0.3.2

Jsonnet v0.11.2

Install ksonnet

Install Kubeflow

After completing the steps, verify the Deployment resources in the cluster and check pod statuses to ensure all services are running.

Use Ambassador to expose Kubeflow components, then forward its service port locally with kubectl port-forward and access the UIs at localhost:8080 .

The UIs let you run Jupyter notebooks for end‑to‑end development and launch TensorFlow operators for multi‑node, multi‑GPU training.

Conclusion

Major cloud providers and hardware vendors are investing in Kubeflow to enable large‑scale, multi‑GPU training on Kubernetes, improving GPU utilization and streamlining the entire ML workflow. This reduces the operational burden on data scientists, letting them focus on algorithms, while presenting new challenges for DevOps teams.

Further exploration of the many components mentioned is encouraged.

References

http://stevenwhang.com/tfx_paper.pdf

https://www.kubeflow.org/

https://opensource.com/article/18/6/kubeflow

https://www.oliverwyman.com/content/dam/oliver-wyman/v2/events/2018/March/Google_London_Event/Public%20Introduction%20to%20Kubeflow.pdf