k8s

Overview ¶

Package k8s provides Kubernetes integration for Cloud Native Stack.

This package contains sub-packages for Kubernetes cluster interaction:

Sub-packages ¶

client: Singleton Kubernetes client with automatic authentication

clientset, config, err := client.GetKubeClient()
if err != nil {
    return err
}
// Use clientset for API operations

agent: Kubernetes Job deployment for automated snapshot capture

deployer := agent.NewDeployer(clientset, agentConfig)
if err := deployer.Deploy(ctx); err != nil {
    return err
}

Architecture ¶

The k8s package follows these design principles:

• Singleton Pattern: The client package uses sync.Once to ensure a single Kubernetes client instance is shared across the application, preventing connection exhaustion and reducing API server load.

• Automatic Authentication: The client automatically detects whether it's running in-cluster (using service account) or out-of-cluster (using kubeconfig file).

• Job-based Agent: The agent package deploys snapshot capture as a Kubernetes Job, enabling GPU node targeting and ConfigMap-based output storage.

Usage Patterns ¶

For most use cases, import and use the client sub-package:

import "github.com/NVIDIA/aicr/pkg/k8s/client"

// Get shared client instance
clientset, _, err := client.GetKubeClient()

For agent deployment, import the agent sub-package:

import "github.com/NVIDIA/aicr/pkg/k8s/agent"

// Deploy snapshot agent
config := agent.Config{
    Namespace: "gpu-operator",
    Image:     "ghcr.io/nvidia/aicr-validator:latest",
}
deployer := agent.NewDeployer(clientset, config)

Thread Safety ¶

Both sub-packages are designed for concurrent use:

• client: Uses sync.Once for thread-safe initialization
• agent: Each Deployer instance is independent