Skip to main content
Katib is Kubeflow’s Kubernetes-native hyperparameter tuning and experiment management system. To install Katib as a standalone deployment, apply the official Kubernetes manifests from the Katib repository. These manifests create controllers, database components, services, and the Katib UI necessary to run experiments inside your cluster.

​
Quick install

Run the following command to deploy Katib (uses the v0.17.0 manifests in this example):
This deploys a complete experimentation platform that manages Experiments, Suggestions, and Trials for hyperparameter optimization.

​
Verify the installation

Katib components are deployed into the kubeflow namespace. Kubernetes items can initialize asynchronously, so allow a few moments for pods to reach the Running state. Check pod status with:
Example output while components are initializing:
Want to watch status updates in real time? Use kubectl get pods -n kubeflow -w to stream pod status changes until they become Ready.
A successful Katib installation typically includes these primary components:
A slide titled "Expected Katib Components" showing four numbered boxes for katib-controller, katib-db-manager, katib-mysql, and katib-ui. Each box includes a brief role: reconciles experiments, manages trial database, stores experiment data, and web dashboard.
Together, these components form the Katib experimentation platform. The Katib UI provides a visual way to monitor experiments, inspect trial runs, compare results, and manage optimization workflows.

​
Accessing the Katib UI

To view the Katib dashboard locally, port-forward the UI service and open the forwarded port in your browser. First, confirm the service name and port:
Then create a port-forward to the Katib UI service (example):
Open http://localhost:8080 (or the local port you forwarded) to access the dashboard. If the service uses a different port, adjust the 8080:80 mapping accordingly.

​
Common issues & troubleshooting

If pods remain in ContainerCreating or CrashLoopBackOff, check container image pulls, node resource availability, and persistent volume claims. Use kubectl describe pod <pod-name> -n kubeflow and kubectl logs <pod-name> -n kubeflow to gather troubleshooting details.
Useful commands for debugging:
  • View events and pod details:
  • Inspect logs of a failing pod:
  • Confirm CRDs and RBAC (Katib requires its CRDs to be present):
By following these steps you will have Katib installed and available for running hyperparameter tuning experiments on your Kubernetes cluster.

Watch Video