Concepts

VKS (Virtual Kubernetes Service) is the GPU-cloud elastic Kubernetes (virtual Kubernetes) cluster offering. Before using it, you'll want to be familiar with the terms below.

Cluster

A Kubernetes cluster is a set of physical or virtual machines, called Nodes. There are two kinds of nodes: a Master Node, which schedules and manages the cluster, and Worker Nodes, which run application containers.

Primary account

The account used to register the company is by default the company administrator (the "primary account"). Each company has exactly one primary account, which holds the highest privileges and can:

• Grant cluster access to other accounts in the company.
• Manage member roles and permissions.
• View and modify company information.
• Manage company-wide resources and service settings.

Parallelism

For a company, the parallelism limit is the upper bound on the total number of GPUs the company may provision. For a single VKS, it is the upper bound on the number of GPUs of a particular type the user may use in that cluster.

Node

A Node is a single machine (physical or virtual) in a Kubernetes cluster. Each node runs the services required to host Pods and to communicate with the Master. Key components on the node include kubelet, kube-proxy, and a container engine (e.g. Docker).

Container

A container packages an application together with its runtime dependencies and is the unit in which applications run. A node can run many containers, typically inside a Pod.

Pod

The Pod is the smallest deployable unit in a cluster. It can contain one or more closely related containers, all sharing the same network namespace, IP, and volumes. A Pod represents a single running process in the cluster.

• Inspect Pod status in a namespace:

kubectl get pod -n [deepseek]

• Open a shell inside a Pod:

kubectl exec -it [prepare-deploy-74f545496-9bl62] -n [deepseek] -- bash

ReplicaSet

ReplicaSet ensures a specified number of Pod replicas are running at any time. If a Pod fails, the ReplicaSet creates a replacement, providing application availability and reliability.

Workload

Service

A Service is an abstraction that defines a logical set of Pods and a policy for accessing them. It provides a stable IP and DNS name so other services can find and reach the Pods. Common types are ClusterIP (in-cluster only), NodePort (a port exposed on each node), and LoadBalancer (uses the cloud provider's load balancer).

Namespace

Namespaces partition resources within a single physical cluster. They help organize objects and isolate them by team or project. Kubernetes ships with three: default, kube-system (system components), and kube-public (public information). Example to create a namespace in VKS:

kubectl create namespace [deepseek]

ConfigMap

Stores configuration data, which can be injected into containers as environment variables, command-line arguments, or configuration files.

Secret

Stores sensitive information such as passwords or OAuth tokens; encrypted at rest by default. Example to create a Secret in VKS:

kubectl apply -f [deepseek-secret.yaml]

Volume

A Volume is a form of persistent storage that survives Pod restarts. Kubernetes supports many volume types, including local storage, network storage (NFS, iSCSI), and cloud storage (AWS EBS, GCE PD).

Ingress

An API object that manages HTTP(S) access rules to services in the cluster. Typically used for host-based routing and SSL termination.

Service Export

Where NodePort cannot be used directly to expose a service, VKS provides a ServiceExporter. Bind a ServiceExporter to a Service that needs external exposure, and it provides a public-facing address.

• Create a ServiceExport resource:

kubectl apply -f [ray-svcExporter-chat.yaml]

• Inspect after creation:

kubectl describe serviceExporter [ray-svc-chat-exporter] -n [deepseek]

Tip: Replace the parameters in [ ] above with the actual resource names you need.