# Copyright 2026 NVIDIA CORPORATION
# SPDX-License-Identifier: Apache-2.0
#
# Kind cluster for the NVSentinel thermal-margin demo.
#
# One control-plane + two workers. The mock GPUs and GPU Operator operands are
# pinned to the workers (run.sh labels them nvml-mock-gpu=true), while the
# NVSentinel control-plane pipeline (MongoDB, platform-connector, fault-quarantine,
# node-drainer) is pinned to the control-plane. When NVSentinel remediates an
# a thermal-margin violation it cordons and drains the *faulty* GPU worker; a GPU workload can
# then reschedule onto the second, healthy worker.
#
# Default kind node naming is used (no kubeadm name overrides) so each node's
# Kubernetes name matches its docker container name — run.sh uses the same names
# for `kubectl` and for `docker exec` (nvidia-container-toolkit install).
#
# CDI is enabled in containerd and the nvidia runtime handler is registered on
# every node; the nvidia-container-toolkit package is installed into each worker
# after creation and containerd restarted (see run.sh).
kind: Cluster
apiVersion: kind.x-k8s.io/v1alpha4
containerdConfigPatches:
  - |-
    [plugins."io.containerd.grpc.v1.cri"]
      enable_cdi = true
    [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia]
      runtime_type = "io.containerd.runc.v2"
    [plugins."io.containerd.grpc.v1.cri".containerd.runtimes.nvidia.options]
      BinaryName = "/usr/bin/nvidia-container-runtime"
nodes:
  - role: control-plane
  - role: worker
  - role: worker