JARK Stack Architecture

JARK Stack Architecture An architecture diagram generated by Archify. Data Scientists & ML Engineers · Notebook users · Architecture component Data Scientists & ML Engineers Notebook users JupyterHub Hub Server · KubeSpawner · OAuth/Cognito · JupyterHub JupyterHub Hub Server KubeSpawner · OAuth/Cognito User Notebooks · CPU · GPU · multi-GPU profiles · JupyterHub User Notebooks CPU · GPU · multi-GPU profiles Workflow Controller · controller separate from server · Argo Workflows Workflow Controller controller separate from server ML Workflows · Prep · Train · Tune · Evaluate · Argo Workflows ML Workflows Prep · Train · Tune · Evaluate Ray Head Node · GCS / dashboard / raylet · Ray Cluster Ray Head Node GCS / dashboard / raylet Ray Worker Pools · CPU · GPU · Neuron · Ray Cluster Ray Worker Pools CPU · GPU · Neuron Karpenter · Node provisioner · Karpenter Auto-Scaling Karpenter Node provisioner Node provisioning policy · NodePool / NodeClass · Karpenter Auto-Scaling Node provisioning policy NodePool / NodeClass Amazon EFS · Notebooks & models · Shared Storage Amazon EFS Notebooks & models FSx for Lustre · Training data · Shared Storage FSx for Lustre Training data Amazon S3 · Datasets & artifacts · Shared Storage Amazon S3 Datasets & artifacts sign in spawn notebook submit workflow run Ray job run training / tuning Pending Pod signal provision JupyterHub Argo Workflows Ray Cluster Karpenter Auto-Scaling Shared Storage Legend Frontend Backend Database Cloud External

Separate control loops

  • • Ray workers do not directly call Karpenter APIs.
  • • Scope user authentication and workflow execution identities separately.
  • • Validate namespace, mount and authorization conditions for shared storage.