Kubeflow

Rather than one application, Kubeflow is a family of Kubernetes-native subprojects that each cover a stage of the ML lifecycle and can be installed alone or together. The graduated set includes Pipelines for containerized workflows, Trainer for distributed training jobs, Katib for hyperparameter tuning, Notebooks for in-cluster development environments, the Spark Operator, and Hub, the former Model Registry, which now adds a catalog of external models. Kale and a unified Python SDK are incubating, while an MLflow integration and an MCP server are experimental. KServe, which began inside Kubeflow as KFServing, is the usual serving layer. The kubeflow/kubeflow repo itself just points to those subprojects. Running the full stack means installing the vendor-neutral Kubeflow Community Distribution, or a packaged one from vendors such as Canonical, onto a cluster with room for many controllers, so expect real cluster administration. Everything is Apache-2.0. Kubeflow began as an open take on Google's internal TensorFlow Extended setup and is now a CNCF graduated project. Suited to platform teams building a shared ML platform.