In this tutorial, we'll network boot and provision a Kubernetes v1.9.3 cluster on bare-metal.

First, we'll deploy a Matchbox service and setup a network boot environment. Then, we'll declare a Kubernetes cluster in Terraform using the Typhoon Terraform module and power on machines. On PXE boot, machines will install Container Linux to disk, reboot into the disk install, and provision themselves as Kubernetes controllers or workers.

Controllers are provisioned as etcd peers and run etcd-member (etcd3) and kubelet. Workers are provisioned to run a kubelet. A one-time bootkube bootstrap schedules an apiserver, scheduler, controller-manager, and kube-dns on controllers and runs kube-proxy and calico or flannel on each node. A generated kubeconfig provides kubectl access to the cluster.


  • Machines with 2GB RAM, 30GB disk, PXE-enabled NIC, IPMI
  • PXE-enabled network boot environment
  • Matchbox v0.6+ deployment with API enabled
  • Matchbox credentials client.crt, client.key, ca.crt
  • Terraform v0.11.x and terraform-provider-matchbox installed locally


Collect a MAC address from each machine. For machines with multiple PXE-enabled NICs, pick one of the MAC addresses. MAC addresses will be used to match machines to profiles during network boot.

  • 52:54:00:a1:9c:ae (node1)
  • 52:54:00:b2:2f:86 (node2)
  • 52:54:00:c3:61:77 (node3)

Configure each machine to boot from the disk 1 through IPMI or the BIOS menu.

ipmitool -H node1 -U USER -P PASS chassis bootdev disk options=persistent

During provisioning, you'll explicitly set the boot device to pxe for the next boot only. Machines will install (overwrite) the operating system to disk on PXE boot and reboot into the disk install.

Ask your hardware vendor to provide MACs and preconfigure IPMI, if possible. With it, you can rack new servers, terraform apply with new info, and power on machines that network boot and provision into clusters.


Create a DNS A (or AAAA) record for each node's default interface. Create a record that resolves to each controller node (or re-use the node record if there's one controller).

  • (node1)
  • (node2)
  • (node3)
  • (node1)

Cluster nodes will be configured to refer to the control plane and themselves by these fully qualified names and they'll be used in generated TLS certificates.


Matchbox is an open-source app that matches network-booted bare-metal machines (based on labels like MAC, UUID, etc.) to profiles to automate cluster provisioning.

Install Matchbox on a Kubernetes cluster or dedicated server.


Deploy Matchbox as service that can be accessed by all of your bare-metal machines globally. This provides a single endpoint to use Terraform to manage bare-metal clusters at different sites. Typhoon will never include secrets in provisioning user-data so you may even deploy matchbox publicly.

Matchbox provides a TLS client-authenticated API that clients, like Terraform, can use to manage machine matching and profiles. Think of it like a cloud provider API, but for creating bare-metal instances.

Generate TLS client credentials. Save the ca.crt, client.crt, and client.key where they can be referenced in Terraform configs.

mv ca.crt client.crt client.key ~/.config/matchbox/

Verify the matchbox read-only HTTP endpoints are accessible (port is configurable).

$ curl

Verify your TLS client certificate and key can be used to access the Matchbox API (port is configurable).

$ openssl s_client -connect \
  -CAfile ~/.config/matchbox/ca.crt \
  -cert ~/.config/matchbox/client.crt \
  -key ~/.config/matchbox/client.key

PXE Environment

Create a iPXE-enabled network boot environment. Configure PXE clients to chainload iPXE and instruct iPXE clients to chainload from your Matchbox service's /boot.ipxe endpoint.

For networks already supporting iPXE clients, you can add a default.ipxe config.

# /var/www/html/ipxe/default.ipxe

For networks with Ubiquiti Routers, you can configure the router itself to chainload machines to iPXE and Matchbox.

For a small lab, you may wish to checkout the container image and copy-paste examples.

Read about the many ways to setup a compliant iPXE-enabled network. There is quite a bit of flexibility:

  • Continue using existing DHCP, TFTP, or DNS services
  • Configure specific machines, subnets, or architectures to chainload from Matchbox
  • Place Matchbox behind a menu entry (timeout and default to Matchbox)

TFTP chainloading to modern boot firmware, like iPXE, avoids issues with old NICs and allows faster transfer protocols like HTTP to be used.

Terraform Setup

Install Terraform v0.11.x on your system.

$ terraform version
Terraform v0.11.1

Add the terraform-provider-matchbox plugin binary for your system.

tar xzf terraform-provider-matchbox-v0.2.2-linux-amd64.tar.gz
sudo mv terraform-provider-matchbox-v0.2.2-linux-amd64/terraform-provider-matchbox /usr/local/bin/

Add the plugin to your ~/.terraformrc.

providers {
  matchbox = "/usr/local/bin/terraform-provider-matchbox"

Read concepts to learn about Terraform, modules, and organizing resources. Change to your infrastructure repository (e.g. infra).

cd infra/clusters


Configure the Matchbox provider to use your Matchbox API endpoint and client certificate in a file.

provider "matchbox" {
  endpoint    = ""
  client_cert = "${file("~/.config/matchbox/client.crt")}"
  client_key  = "${file("~/.config/matchbox/client.key")}"
  ca          = "${file("~/.config/matchbox/ca.crt")}"

provider "local" {
  version = "~> 1.0"
  alias = "default"

provider "null" {
  version = "~> 1.0"
  alias = "default"

provider "template" {
  version = "~> 1.0"
  alias = "default"

provider "tls" {
  version = "~> 1.0"
  alias = "default"


Define a Kubernetes cluster using the module bare-metal/container-linux/kubernetes.

module "bare-metal-mercury" {
  source = "git::"

  providers = {
    local = "local.default"
    null = "null.default"
    template = "template.default"
    tls = "tls.default"

  # install
  matchbox_http_endpoint  = ""
  container_linux_channel = "stable"
  container_linux_version = "1576.5.0"
  ssh_authorized_key      = "ssh-rsa AAAAB3Nz..."

  # cluster
  cluster_name    = "mercury"
  k8s_domain_name = ""

  # machines
  controller_names   = ["node1"]
  controller_macs    = ["52:54:00:a1:9c:ae"]
  controller_domains = [""]
  worker_names = [
  worker_macs = [
  worker_domains = [

  # output assets dir
  asset_dir = "/home/user/.secrets/clusters/mercury"

Reference the variables docs or the source.


Initial bootstrapping requires bootkube.service be started on one controller node. Terraform uses ssh-agent to automate this step. Add your SSH private key to ssh-agent.

ssh-add ~/.ssh/id_rsa
ssh-add -L


terraform apply will hang connecting to a controller if ssh-agent does not contain the SSH key.


Initialize the config directory if this is the first use with Terraform.

terraform init

Get or update Terraform modules.

$ terraform get            # downloads missing modules
$ terraform get --update   # updates all modules
Get: git:: (update)
Get: git:: (update)

Plan the resources to be created.

$ terraform plan
Plan: 55 to add, 0 to change, 0 to destroy.

Apply the changes. Terraform will generate bootkube assets to asset_dir and create Matchbox profiles (e.g. controller, worker) and matching rules via the Matchbox API.

module.bare-metal-mercury.null_resource.copy-kubeconfig.0: Provisioning with 'file'...
module.bare-metal-mercury.null_resource.copy-etcd-secrets.0: Provisioning with 'file'...
module.bare-metal-mercury.null_resource.copy-kubeconfig.0: Still creating... (10s elapsed)
module.bare-metal-mercury.null_resource.copy-etcd-secrets.0: Still creating... (10s elapsed)

Apply will then loop until it can successfully copy credentials to each machine and start the one-time Kubernetes bootstrap service. Proceed to the next step while this loops.


Power on each machine with the boot device set to pxe for the next boot only.

ipmitool -H -U USER -P PASS chassis bootdev pxe
ipmitool -H -U USER -P PASS power on

Machines will network boot, install Container Linux to disk, reboot into the disk install, and provision themselves as controllers or workers.

If this is the first test of your PXE-enabled network boot environment, watch the SOL console of a machine to spot any misconfigurations.


Wait for the bootkube-start step to finish bootstrapping the Kubernetes control plane. This may take 5-15 minutes depending on your network.

module.bare-metal-mercury.null_resource.bootkube-start: Still creating... (6m10s elapsed)
module.bare-metal-mercury.null_resource.bootkube-start: Still creating... (6m20s elapsed)
module.bare-metal-mercury.null_resource.bootkube-start: Still creating... (6m30s elapsed)
module.bare-metal-mercury.null_resource.bootkube-start: Still creating... (6m40s elapsed)
module.bare-metal-mercury.null_resource.bootkube-start: Creation complete (ID: 5441741360626669024)

Apply complete! Resources: 55 added, 0 changed, 0 destroyed.

To watch the bootstrap process in detail, SSH to the first controller and journal the logs.

$ ssh
$ journalctl -f -u bootkube
bootkube[5]:         Pod Status:        pod-checkpointer        Running
bootkube[5]:         Pod Status:          kube-apiserver        Running
bootkube[5]:         Pod Status:          kube-scheduler        Running
bootkube[5]:         Pod Status: kube-controller-manager        Running
bootkube[5]: All self-hosted control plane components successfully started
bootkube[5]: Tearing down temporary bootstrap control plane...


Install kubectl on your system. Use the generated kubeconfig credentials to access the Kubernetes cluster and list nodes.

$ export KUBECONFIG=/home/user/.secrets/clusters/mercury/auth/kubeconfig
$ kubectl get nodes
NAME                STATUS    AGE       VERSION   Ready     11m       v1.9.3   Ready     11m       v1.9.3   Ready     11m       v1.9.3

List the pods.

$ kubectl get pods --all-namespaces
NAMESPACE     NAME                                       READY     STATUS    RESTARTS   AGE
kube-system   calico-node-6qp7f                          2/2       Running   1          11m
kube-system   calico-node-gnjrm                          2/2       Running   0          11m
kube-system   calico-node-llbgt                          2/2       Running   0          11m
kube-system   kube-apiserver-7336w                       1/1       Running   0          11m
kube-system   kube-controller-manager-3271970485-b9chx   1/1       Running   0          11m
kube-system   kube-controller-manager-3271970485-v30js   1/1       Running   1          11m
kube-system   kube-dns-1187388186-mx9rt                  3/3       Running   0          11m
kube-system   kube-proxy-50sd4                           1/1       Running   0          11m
kube-system   kube-proxy-bczhp                           1/1       Running   0          11m
kube-system   kube-proxy-mp2fw                           1/1       Running   0          11m
kube-system   kube-scheduler-3895335239-fd3l7            1/1       Running   1          11m
kube-system   kube-scheduler-3895335239-hfjv0            1/1       Running   0          11m
kube-system   pod-checkpointer-wf65d                     1/1       Running   0          11m
kube-system   1/1       Running   0          11m

Going Further

Learn about version pinning, maintenance, and addons.


On Container Linux clusters, install the container-linux-update-operator addon to coordinate reboots and drains when nodes auto-update. Otherwise, updates may not be applied until the next reboot.



Name Description Example
matchbox_http_endpoint Matchbox HTTP read-only endpoint
container_linux_channel Container Linux channel stable, beta, alpha
container_linux_version Container Linux version of the kernel/initrd to PXE and the image to install 1576.5.0
cluster_name Cluster name mercury
k8s_domain_name FQDN resolving to the controller(s) nodes. Workers and kubectl will communicate with this endpoint ""
ssh_authorized_key SSH public key for ~/.ssh/authorized_keys "ssh-rsa AAAAB3Nz..."
controller_names Ordered list of controller short names ["node1"]
controller_macs Ordered list of controller identifying MAC addresses ["52:54:00:a1:9c:ae"]
controller_domains Ordered list of controller FQDNs [""]
worker_names Ordered list of worker short names ["node2", "node3"]
worker_macs Ordered list of worker identifying MAC addresses ["52:54:00:b2:2f:86", "52:54:00:c3:61:77"]
worker_domains Ordered list of worker FQDNs ["", ""]
asset_dir Path to a directory where generated assets should be placed (contains secrets) "/home/user/.secrets/clusters/mercury"


Name Description Default Example
cached_install Whether machines should PXE boot and install from the Matchbox /assets cache. Admin MUST have downloaded Container Linux images into the cache to use this false true
install_disk Disk device where Container Linux should be installed "/dev/sda" "/dev/sdb"
container_linux_oem Specify alternative OEM image ids for the disk install "" "vmware_raw", "xen"
networking Choice of networking provider "calico" "calico" or "flannel"
network_mtu CNI interface MTU (calico-only) 1480 -
pod_cidr CIDR range to assign to Kubernetes pods "" ""
service_cidr CIDR range to assign to Kubernetes services "" ""
cluster_domain_suffix FQDN suffix for Kubernetes services answered by kube-dns. "cluster.local" ""
kernel_args Additional kernel args to provide at PXE boot [] "kvm-intel.nested=1"

  1. Configuring "diskless" workers that always PXE boot is possible, but not in the scope of this tutorial.