You are viewing documentation for Kubeflow 1.2

This is a static snapshot from the time of the Kubeflow 1.2 release.
For up-to-date information, see the latest version.

Resuming an Experiment

How to restart and modify running experiments

This guide describes how to modify running experiments and restart completed experiments. You will learn about changing the experiment execution process and use various resume policies for the Katib experiment.

For the details on how to configure and run your experiment, follow the running an experiment guide.

Modify running experiment

While the experiment is running you are able to change trial count parameters. For example, you can decrease the maximum number of hyperparameter sets that are trained in parallel.

You can change only parallelTrialCount, maxTrialCount and maxFailedTrialCount experiment parameters.

Use Kubernetes API or kubectl in-place update of resources to make experiment changes. For example, run:

kubectl edit experiment <experiment-name> -n <experiment-namespace>

Make appropriate changes and save it. Controller automatically processes the new parameters and makes necessary changes.

  • If you want to increase or decrease parallel trial execution, modify parallelTrialCount. Controller accordingly creates or deletes trials in line with the parallelTrialCount value.

  • If you want to increase or decrease maximum trial count, modify maxTrialCount. maxTrialCount should be greater than current count of Succeeded trials. You can remove the maxTrialCount parameter, if your experiment should run endless with parallelTrialCount of parallel trials until the experiment reaches Goal or maxFailedTrialCount

  • If you want to increase or decrease maximum failed trial count, modify maxFailedTrialCount. You can remove the maxFailedTrialCount parameter, if the experiment should not reach Failed status.

Resume succeeded experiment

Katib experiment is restartable only if it is in Succeeded status because maxTrialCount has been reached. To check current experiment status run: kubectl get experiment <experiment-name> -n <experiment-namespace>.

To restart an experiment, you are able to change only parallelTrialCount, maxTrialCount and maxFailedTrialCount as described above

To control various resume policies, you can specify .spec.resumePolicy for the experiment. Refer to the ResumePolicy type.

Resume policy: Never

Use this policy if your experiment should not be resumed at any time. After the experiment has finished, the suggestion’s Deployment and Service are deleted and you can’t restart the experiment. Learn more about Katib concepts in the overview guide.

Check the never-resume.yaml example for more details.

Resume policy: LongRunning

Use this policy if you intend to restart the experiment. After the experiment has finished, the suggestion’s Deployment and Service stay running. Modify experiment’s trial count parameters to restart the experiment.

When you delete the experiment, the suggestion’s Deployment and Service are deleted.

This is the default policy for all Katib experiments. You can omit .spec.resumePolicy parameter for that functionality.

Resume policy: FromVolume

Use this policy if you intend to restart the experiment. In that case, volume is attached to the suggestion’s Deployment.

Katib controller creates PersistentVolumeClaim (PVC) and PersistentVolume (PV) in addition to the suggestion’s Deployment and Service.

Note: If you specify storage class name in the Katib configuration settings for the PVC differently from the katib-suggestion, Katib controller creates only PVC. Follow the Katib configuration guide to set up the suggestion’s volume settings.

  • PVC is deployed with the name: <suggestion-name>-<suggestion-algorithm> in the suggestion namespace.

  • PV is deployed with the name: <suggestion-name>-<suggestion-algorithm>-<suggestion-namespace>

After the experiment has finished, the suggestion’s Deployment and Service are deleted. Suggestion data can be retained in the volume. When you restart the experiment, the suggestion’s Deployment and Service are created and suggestion statistics can be recovered from the volume.

When you delete the experiment, the suggestion’s Deployment, Service, PVC and PV are deleted automatically.

Check the from-volume-resume.yaml example for more details.

Next steps