Skip to content
Search

Kubeflow:Architecture

lastmod 2026-03-01created 2025-12-2925 edits

This guide introduces the Kubeflow projects and how they are used in each stage of the AI lifecycle.

Kubeflow Ecosystem

The following diagram gives an overview of the Kubeflow ecosystem and shows how it relates to the wider Kubernetes and AI landscape. Kubeflow is built on Kubernetes as a system for deploying, scaling, and managing AI platforms.

Kubeflow_Ecosystem.png

Introducing the AI Lifecycle

When developing and deploying artificial intelligence (AI) applications, the AI lifecycle typically consists of several stages. Developing an AI system is an iterative process. You need to evaluate the outputs produced at the various stages of the AI lifecycle and modify the model and parameters as needed so that the model keeps producing the required results.

The following diagram shows the stages of the AI lifecycle in order.

Introducing_the_AI_Lifecycle.png

Let's look at each stage in more detail.

  • Data Preparation stage - collect raw data, perform feature engineering to extract machine learning features to store in an offline feature store, and prepare training data for model development.
    • Typically this stage involves data processing tools such as Spark, Dask, Flink, or Ray.
  • Model Development stage - choose a machine learning framework, develop the model architecture, and explore existing pre-trained models such as BERT or Llama to fine-tune them.
  • Model Training stage - train or fine-tune models in a large-scale compute environment.
    • If a single GPU cannot handle the model size, you need to use distributed training.
    • The output of model training is a trained model artifact, which can be stored in the model registry.
  • Model Optimization stage - optimize model hyperparameters and optimize models using various AutoML algorithms such as neural architecture search and model compression.
    • During model optimization, machine learning metadata can be stored in the model registry.
  • Model Serving stage - serve model artifacts for online or batch inference.
    • Depending on the use case, the model can perform predictive or generative AI tasks.
    • In the model serving stage, features can be extracted using an online feature store.
    • Model performance is monitored and the results are fed back into earlier stages of the AI lifecycle.

AI Lifecycle for Production and Development Phases

The AI lifecycle of AI applications can conceptually be divided into a development phase and a production phase, and this diagram shows which tasks each phase includes.

AI_Lifecycle_for_Production_and_Development_Phases.png

Kubeflow Projects in the AI Lifecycle

The following diagram shows how the Kubeflow projects fit into each stage of the AI lifecycle.

Kubeflow_Projects_in_the_AI_Lifecycle.png

See the following links for details on each Kubeflow project.

  • Kubeflow Spark Operator can be used for the data preparation and feature engineering stages.
  • Kubeflow Notebooks can be used for model development and interactive data science to experiment with AI workflows.
  • Kubeflow Trainer can be used for large-scale distributed training or LLM (Learning Leadership Model) fine-tuning.
  • Kubeflow Katib can be used for model optimization and hyperparameter tuning with various AutoML algorithms.
  • Kubeflow Model Registry can be used to store machine learning metadata and model artifacts and to prepare models for deployment to production environments.
  • KServe can be used for both online and batch inference in the model serving stage.
  • Feast can be used as a feature store and can manage offline and online features.
  • Kubeflow Pipelines can be used to build, deploy, and manage each stage of the AI lifecycle.

AI platform teams can build on Kubeflow to meet their specific requirements by using each project independently or by deploying the full AI reference platform

Kubeflow APIs and SDKs

Various Kubeflow projects provide APIs and Python SDKs.

See the following reference documentation.

What if we turn it into a gRPC API?

Cluster control:

S3 compatibility layer:

Mesh network:

Authentication/user management/security:

  • Cert Manager <- domain registration (?)
  • Dex
  • OAuth2-Proxy <- needs to integrate with the security system already in use.
  • Profiles + KFAM

Filesystem/Data Access:

MLOps Pipeline:

Other:

  • Central Dashboard
  • PodDefaults Webhook

Timeline

  • January 2026
    • Cluster control
    • Kubernetes REST API -> gRPC Interface Wrapping
    • Network Interface API -> exposing endpoints using Traefik/Envoy
  • February 2026
    • Filesystem/Data Access API
    • Kubeflow PVC Viewer Controller API
    • Filesystem API / CRUD
  • March 2026
    • Apply SeaweedFS as the S3 compatibility layer
    • Kubeflow Spark Operator

See also

Favorite site

author Zer0 sitemap