Polyaxon provides native support for several KubeFlow components.
Overview
- Kubeflow Operators:
Polyaxon can schedule and manage Kubeflow operators natively. Polyaxon provides a uniform workflow for:
- Viewing logs and resources.
- Tracking metrics, outputs, and models.
- Comparing and driving insights.
- Kubeflow Pipelines: Polyaxon supports Kubeflow Pipeline components with very few changes.
- Kubeflow KFServing: Polyaxon provides reusable components that can deploy models using KFServing.
Deploying Kubeflow’s training jobs operator
For teams not running/using Kubeflow and want to use this integration, Polyaxon provides a Helm chart for the Kubeflow operators currently supported.
The Helm chart will be maintained and supported by Polyaxon to allow users to deploy and manage Kubeflow Training Jobs Operator in an easy way.
This operator requires Helm to be installed.
We are also distributing the chart directly on our official Helm charts repo https://charts.polyaxon.com
helm repo add polyaxon https://charts.polyaxon.com
helm repo update
Deploying/Deleting the TrainingJobs operator
In order to use Kubeflow as a backend for running:
- distributed Tensorflow experiments
- distributed Pytorch experiments
- distributed MPI experiments
- distributed MXNet experiments
- distributed XGBoost experiments
you need to deploy polyaxon/trainingjobs
on the same namespace where Polyaxon (CE or Agent) is deployed
helm install trainingjobs polyaxon/trainingjobs --namespace=polyaxon
helm del trainingjobs --purge