Skip to main content

Self-hosted runner gVisor setup

Why a self-hosted runner?​

The gvisor-smoke.yml CI workflow is workflow_dispatch-only rather than running on every push. The reason: GitHub-hosted runners use Docker to run Kind node containers, and Docker containers apply capability restrictions that prevent gVisor's runsc from launching sandboxed containers. Pods using runtimeClassName: gvisor inside a Kind cluster on a GitHub-hosted runner stay in ContainerCreating indefinitely.

The Kubernetes-level tests (RuntimeClass wiring, pod spec validation) DO pass because they only inspect API object specs. The exec-based compatibility tests (git clone, curl, shell commands) require a pod that actually runs under gVisor, which needs either:

  1. A self-hosted runner with gVisor installed on the bare metal / VM host (not nested inside Docker) — instructions below.
  2. A managed Kubernetes cluster with a gVisor node pool (GKE Sandbox, EKS Bottlerocket with gVisor) — run the test suite directly against that cluster without Kind at all.

Setting up a self-hosted runner with gVisor​

Prerequisites​

RequirementNotes
Ubuntu 22.04 LTS host (bare metal or VM)Not inside Docker
Kernel 5.15+Recommended; gVisor works from 4.14 but 5.15+ has better syscall coverage
containerd 1.7+Not Docker-only — containerd as a standalone daemon
GitHub Actions runner binaryv2.300+ recommended

Step 1: Install gVisor​

# Add the gVisor apt repository.
curl -fsSL https://gvisor.dev/archive.key \
  | sudo gpg --dearmor -o /usr/share/keyrings/gvisor-archive-keyring.gpg

echo "deb [arch=$(dpkg --print-architecture) \
  signed-by=/usr/share/keyrings/gvisor-archive-keyring.gpg] \
  https://storage.googleapis.com/gvisor/releases release main" \
  | sudo tee /etc/apt/sources.list.d/gvisor.list

sudo apt-get update && sudo apt-get install -y runsc

# Verify.
runsc --version
which containerd-shim-runsc-v1

Step 2: Configure containerd​

# Add the gVisor runtime handler to containerd's config.
sudo mkdir -p /etc/containerd/config.d

sudo tee /etc/containerd/config.d/gvisor.toml <<'EOF'
[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runsc]
  runtime_type = "io.containerd.runsc.v1"
[plugins."io.containerd.grpc.v1.cri".containerd.runtimes.runsc.options]
  TypeUrl = "io.containerd.runsc.v1.options"
EOF

sudo systemctl restart containerd
sudo systemctl is-active containerd

Step 3: Install and register the GitHub Actions runner​

mkdir -p ~/actions-runner && cd ~/actions-runner

# Download the runner binary (check for latest version at
# https://github.com/actions/runner/releases).
curl -o actions-runner-linux-x64-2.320.0.tar.gz -L \
  https://github.com/actions/runner/releases/download/v2.320.0/actions-runner-linux-x64-2.320.0.tar.gz
tar xzf ./actions-runner-linux-x64-2.320.0.tar.gz

# Configure the runner. Get the token from:
# Settings > Actions > Runners > New self-hosted runner > Linux.
./config.sh \
  --url https://github.com/olafkfreund/AIFactory \
  --token <YOUR_RUNNER_REGISTRATION_TOKEN> \
  --labels gvisor,linux,x64 \
  --name aifactory-gvisor-runner-01

# Install as a systemd service so it restarts automatically.
sudo ./svc.sh install
sudo ./svc.sh start
sudo ./svc.sh status

Step 4: Update the workflow to target the runner​

Modify .github/workflows/gvisor-smoke.yml to use the self-hosted runner:

jobs:
  gvisor-smoke:
    name: gVisor live-cluster smoke (Kind + runsc)
    runs-on: [self-hosted, gvisor, linux, x64]  # was: ubuntu-22.04

Also add the push trigger back:

on:
  push:
    branches: [dev, main]
  pull_request:
    branches: [dev]
  workflow_dispatch:

Step 5: Verify the setup​

After the runner is registered, trigger the workflow manually:

gh workflow run gvisor-smoke.yml

A successful run proves gVisor works end-to-end:

  • RuntimeClass wiring (chart-level)
  • Live pod scheduling under gVisor
  • git clone, curl HTTPS, coreutils all working under runsc

Alternative: run against a managed cluster​

If you have a cluster with a gVisor node pool (e.g. GKE Sandbox, EKS with Bottlerocket + gVisor), you can run the test suite directly without Kind:

# Point at the real cluster.
export KUBECONFIG=~/.kube/your-cluster.yaml
export GVISOR_NAMESPACE=aifactory
export GVISOR_COMPAT_POD=gvisor-compat-tester

# Deploy AIFactory with gVisor enabled.
helm install aifactory charts/aifactory/ \
  --namespace aifactory --create-namespace \
  --set sandbox.gvisor.enabled=true \
  ...

# Launch a test pod with runtimeClassName=gvisor.
kubectl run gvisor-compat-tester \
  --image=alpine:3.19 \
  --namespace=aifactory \
  --restart=Never \
  --overrides='{"spec":{"runtimeClassName":"gvisor"}}' \
  -- sleep 3600

# Wait for it to reach Running.
kubectl wait pod/gvisor-compat-tester \
  -n aifactory --for=condition=Ready --timeout=120s

# Install tools.
kubectl exec -n aifactory gvisor-compat-tester -- \
  apk add --quiet git curl

# Run the smoke tests.
pytest tests/helm/test_live_gvisor.py -m gvisor_live -v