Requirements

This document outlines the specific requirements for deploying Serverless Pods, based on the demo of the Rafay platform's new compute service.

• Kubernetes Cluster

  • A host Kubernetes cluster must be provisioned and managed by the platform.
  • The cluster must include GPU-enabled worker nodes with specific hardware (e.g., NVIDIA A100s, H100s) to support accelerated workloads.
  • The platform's agent must have appropriate permissions to create and manage pods, services, secrets, and other necessary resources.

• Platform Components

  • Environment Manager (EM) and a dedicated agent are required to orchestrate the provisioning of the serverless pods and related resources.
  • The platform's Inventory API is a crucial component for resource matchmaking. It ensures that sufficient CPU, memory, and GPU resources are available before a deployment is scheduled, preventing pods from entering a pending state.

• Resource Management

  • Resource Quotas: Sufficient CPU and memory must be available in the target namespace. Each pod will have its resource requests and limits set to be equal to guarantee a specific quality of service (QoS).
  • GPU Inventory: The Inventory API must be configured to track the availability of specific GPU types on each node.

• Container Images

  • Pod images (e.g., Ubuntu 22.04) must be pre-configured and accessible from the cluster.

• Network and Security

  • SSH Pods: A NodePort service must be created for each SSH-enabled pod to expose port 22 for user access.
  • SSH Key Management: The platform must automatically generate and securely store SSH key pairs in a Kubernetes secret for user authentication.

---