Kubeflow

Kubeflow is an open-source platform for running AI and machine learning work on Kubernetes. Google started it in 2017 as an open-source version of how it ran TensorFlow internally. It is now a CNCF graduated project run by community working groups. It is built from modular subprojects that you can use on their own or together through the Kubeflow Community Distribution. These include Pipelines for workflow orchestration, Notebooks and Workspaces, Trainer for distributed training, Katib for hyperparameter tuning, Hub as a model registry and catalog, a Spark Operator, a central Dashboard, the Kubeflow SDK and an MCP server. Related projects such as KServe and Feast cover model serving and feature stores. Pipelines exposes a REST API and has a command-line tool for managing pipelines and runs from code. It suits ML platform teams at mid-sized and large organisations that already run Kubernetes and want a portable, vendor-neutral MLOps stack on-premises or on any cloud. The software itself is free. Costs come from the Kubernetes cluster, the GPUs and the engineering time needed to run it. Vendors such as Canonical sell managed or supported distributions, including a Managed Kubeflow offer on the Azure Marketplace. Kubeflow does not host anything for you, and it has no service-level agreement or vendor support unless you buy them through a distribution. Its main limitation is operational complexity. Installing and maintaining it involves Helm charts, Istio and many custom resources. Reviewers point to a steep learning curve, scattered documentation and hard-to-read pipeline errors.