Skip to main content
CKS clusters are suitable for hosting containerized training tasks, inference services, batch processing tasks, or long-running business applications. Without going into specific parameter configurations, this document helps you understand the basic workflow: preparing a business image, creating a cluster, configuring storage, deploying workloads, configuring access endpoints, and monitoring operational status.

​
Usage workflow

1

Prepare a Container Image

Kubernetes workloads are typically delivered with their runtime environments through container images. Before submitting a workload, build or obtain the required image for your application, and push it to a container image registry.After the image is prepared, record the image address and tag required for subsequent workloads. For instructions on pushing and pulling images, see Push and Pull Images.
2

Create a CKS Cluster

Log in to the GPU Cloud console, go to the Clusters page, and create a CKS cluster.When creating a cluster, you typically need to confirm the Kubernetes version, node pool configuration, and other settings. For details, see Create a CKS Cluster.
3

Configure cluster access and basic resources

After the cluster is created, prepare the access credentials and basic resources according to the subsequent operation method.
4

Configure storage and data access

After the cluster becomes available, configure storage and data access based on the workload’s data access method.
  • When multiple nodes or workloads need to share data through a file system, you can use the GPFS file system and make it available to workloads through Kubernetes storage objects such as PVs, PVCs, or StorageClasses. For details, see CKS cluster persistent storage overview.
  • When you need to store datasets, model files, inference inputs, training outputs, or archived logs, and upload or download them from a local environment, CES instances, or automated scripts, you can use OSS. For details, see OSS object storage overview.
If the workload needs to mount a directory directly, prepare persistent storage before deployment. If training data, model weights, or output results need to be retained long term, plan the storage location and access permissions in advance.
5

Submit and run workloads

Once the cluster is available, submit workloads for training, inference, batch processing, or online services based on your business requirements.Before submitting, we recommend confirming the workload type, image address and labels, CPU/memory/GPU resource configuration, namespace, configuration items, storage mounts or object storage access methods, and log output path. For deployment and troubleshooting procedures, see Deploy and Troubleshoot Workloads.
6

Configure Service Access

After the workload is running, you can configure a Service or Ingress based on the access scope.
  • Access endpoints within the cluster are typically provided through Services. For details, see Create and Manage Services.
  • HTTP or HTTPS access endpoints are typically provided through Ingresses. If your workload needs to be accessed from outside the cluster, confirm the access domain name or port, target service, and access endpoint. For details, see Expose Cluster Services Using Ingress.
7

Observe Runtime Status

During workload execution, you can use monitoring to view cluster, node, or workload metrics, and use logs to view task output or application errors. For details, see View Cluster Overview Monitoring and Query and Analyze Logs.To continuously monitor cluster or workload exceptions, you can configure a notifier. For details, see Create a Custom Alert Rule.
8

Clean Up Resources

After the task is complete, clean up workloads, image artifacts, and stored data that are no longer needed based on your business requirements. Before deleting a cluster, repository, or storage resource, ensure that no running tasks will be affected and that any data that needs to be retained has been backed up or migrated.