Kubeflow
chirashi.kubeflow
A Docker Compose file for Kubeflow alone, generated from the catalog and checked by the safety inspector. No account needed.
Machine-learning platform for Kubernetes covering pipelines, distributed training operators, hyperparameter search, model registry and multi-user notebook servers under namespace isolation.
Kubeflow is a set of Kubernetes-native components rather than one application: Pipelines compiles a Python-defined workflow into a graph of containerised steps with recorded artifacts, training operators run distributed jobs for the major frameworks, Katib performs hyperparameter search as Kubernetes resources, and the Notebooks controller provisions per-user workspaces. Multi-tenancy is enforced through namespace profiles, keeping each user's workloads, storage and credentials separate. Installation means applying manifests to a cluster and reconciling component versions against the cluster's own; the image referenced here is one notebook workspace container from that catalog rather than the platform.
You know it worked when
- The central dashboard loads and a user profile namespace exists.
- A notebook server provisions and opens in the browser.
- A compiled pipeline is submitted and its steps complete in order.
- Artifacts from that run are recorded and retrievable.
- A second user sees only the resources in their own namespace.
Known sharp edges
- Kubeflow requires a Kubernetes cluster with its controllers and custom resources installed; one container from its catalog provides a notebook workspace and nothing more.
- Component versions are released against specific Kubernetes and service-mesh versions, and a mismatch surfaces as pods stuck pending rather than as a clear error.
- The default installation manifests ship well-known development credentials that must be replaced before anything is exposed.
- Pipeline artifacts and metadata need object storage and a database provisioned separately; without them runs complete but record nothing.