Kubeflow on EKS Deep Dive
Supported Versions: Kubeflow Community Distribution 26.03 Last Updated: August 19, 2026
Overview
Kubeflow is an open-source machine learning platform for Kubernetes that bundles the pieces a team needs to run ML workloads end to end — pipeline orchestration, notebooks, hyperparameter tuning, distributed training, and model serving — as a set of Kubernetes-native controllers and CRDs rather than a single monolithic application. On August 17, 2026, the CNCF announced Kubeflow's graduation (having joined as an incubating project in 2023), following an independent security audit and the formation of a formal steering committee — a strong signal of the project's production maturity.
Component Map
| Component | Problem It Solves | CRD / Core Concept | Deep Dive |
|---|---|---|---|
| Central Dashboard & Profiles | Multi-tenant access, per-user namespace isolation | Profile (namespace) | Part 1 |
| Kubeflow Pipelines | Orchestrate multi-step ML workflows as DAGs | Pipeline, Run, Experiment | Part 2 |
| Kubeflow Notebooks | Managed, per-user Jupyter/RStudio/VS Code environments | Notebook | Part 3 |
| Katib | Hyperparameter tuning and AutoML | Experiment, Trial, Suggestion | Part 4 |
| Kubeflow Trainer | Distributed model training across frameworks | TrainJob, ClusterTrainingRuntime | Part 5 |
| KServe | Model serving and inference | InferenceService | Part 6 |
Why Run This on EKS
Kubeflow's components are designed to run on any conformant Kubernetes cluster, which means the operational practices this docs site already covers for EKS — Karpenter-driven autoscaling (including GPU node pools), IRSA/Pod Identity for AWS service access, EBS/S3 storage integration, and observability with Prometheus/Grafana — apply directly to ML workloads rather than requiring a separate ML-specific platform. The trade-off against fully managed alternatives (e.g. Amazon SageMaker) is the same one covered in Data on EKS: more operational responsibility (Operator upgrades, storage/identity wiring) in exchange for a single deployment/observability model shared across all workloads on the cluster, and the ability to run any of Kubeflow's components independently rather than adopting the whole platform at once.
Currently Covered
- Part 1: Kubeflow Architecture and Installation on EKS — component architecture, CNCF graduation context, installing via
awslabs/kubeflow-manifestson EKS - Part 2: Kubeflow Pipelines — KFP SDK v2, IR-based pipeline compilation, S3-backed artifact storage
- Part 3: Kubeflow Notebooks — per-user notebook servers, Profile-based multi-tenancy, GPU scheduling
- Part 4: Katib — Hyperparameter Tuning and AutoML — Experiment/Trial/Suggestion model, search algorithms, early stopping
- Part 5: Kubeflow Trainer and Distributed Training — the v1 Training Operator to Kubeflow Trainer v2 transition, TrainJob/TrainingRuntime
- Part 6: KServe — Model Serving on Kubernetes — InferenceService, Serverless vs. Raw Deployment mode, canary rollouts