e6976a54e1
* add pre-commit hook configuration * add tmp.md to .gitignore * describe the use of pre-commit hook in CONTRIBUTING.md * fix docs/integration.md errors identified by markdownlint * fix docs/<file>.md errors identified by markdownlint * docs/azure-csi.md * docs/azure.md * docs/bootstrap-os.md * docs/calico.md * docs/debian.md * docs/fcos.md * docs/vagrant.md * docs/gcp-lb.md * docs/kubernetes-apps/registry.md * docs/setting-up-your-first-cluster.md * docs/vagrant.md * docs/vars.md * fix contrib/<file>.md errors identified by markdownlint
164 lines
8.1 KiB
Markdown
164 lines
8.1 KiB
Markdown
# Vagrant
|
|
|
|
Assuming you have Vagrant 2.0+ installed with virtualbox, libvirt/qemu or
|
|
vmware, but is untested) you should be able to launch a 3 node Kubernetes
|
|
cluster by simply running `vagrant up`.
|
|
|
|
This will spin up 3 VMs and install kubernetes on them.
|
|
Once they are completed you can connect to any of them by running `vagrant ssh k8s-[1..3]`.
|
|
|
|
To give an estimate of the expected duration of a provisioning run:
|
|
On a dual core i5-6300u laptop with an SSD, provisioning takes around 13
|
|
to 15 minutes, once the container images and other files are cached.
|
|
Note that libvirt/qemu is recommended over virtualbox as it is quite a bit
|
|
faster, especially during boot-up time.
|
|
|
|
For proper performance a minimum of 12GB RAM is recommended.
|
|
It is possible to run a 3 node cluster on a laptop with 8GB of RAM using
|
|
the default Vagrantfile, provided you have 8GB zram swap configured and
|
|
not much more than a browser and a mail client running.
|
|
If you decide to run on such a machine, then also make sure that any tmpfs
|
|
devices, that are mounted, are mostly empty and disable any swapfiles
|
|
mounted on HDD/SSD or you will be in for some serious swap-madness.
|
|
Things can get a bit sluggish during provisioning, but when that's done,
|
|
the system will actually be able to perform quite well.
|
|
|
|
## Customize Vagrant
|
|
|
|
You can override the default settings in the `Vagrantfile` either by
|
|
directly modifying the `Vagrantfile` or through an override file.
|
|
In the same directory as the `Vagrantfile`, create a folder called
|
|
`vagrant` and create `config.rb` file in it.
|
|
An example of how to configure this file is given below.
|
|
|
|
## Use alternative OS for Vagrant
|
|
|
|
By default, Vagrant uses Ubuntu 18.04 box to provision a local cluster.
|
|
You may use an alternative supported operating system for your local cluster.
|
|
|
|
Customize `$os` variable in `Vagrantfile` or as override, e.g.,:
|
|
|
|
```ShellSession
|
|
echo '$os = "flatcar-stable"' >> vagrant/config.rb
|
|
```
|
|
|
|
The supported operating systems for vagrant are defined in the `SUPPORTED_OS`
|
|
constant in the `Vagrantfile`.
|
|
|
|
## File and image caching
|
|
|
|
Kubespray can take quite a while to start on a laptop. To improve provisioning
|
|
speed, the variable 'download_run_once' is set. This will make kubespray
|
|
download all files and containers just once and then redistributes them to
|
|
the other nodes and as a bonus, also cache all downloads locally and re-use
|
|
them on the next provisioning run. For more information on download settings
|
|
see [download documentation](/docs/downloads.md).
|
|
|
|
## Example use of Vagrant
|
|
|
|
The following is an example of setting up and running kubespray using `vagrant`.
|
|
For repeated runs, you could save the script to a file in the root of the
|
|
kubespray and run it by executing `source <name_of_the_file>`.
|
|
|
|
```ShellSession
|
|
# use virtualenv to install all python requirements
|
|
VENVDIR=venv
|
|
virtualenv --python=/usr/bin/python3.7 $VENVDIR
|
|
source $VENVDIR/bin/activate
|
|
pip install -r requirements.txt
|
|
|
|
# prepare an inventory to test with
|
|
INV=inventory/my_lab
|
|
rm -rf ${INV}.bak &> /dev/null
|
|
mv ${INV} ${INV}.bak &> /dev/null
|
|
cp -a inventory/sample ${INV}
|
|
rm -f ${INV}/hosts.ini
|
|
|
|
# customize the vagrant environment
|
|
mkdir vagrant
|
|
cat << EOF > vagrant/config.rb
|
|
\$instance_name_prefix = "kub"
|
|
\$vm_cpus = 1
|
|
\$num_instances = 3
|
|
\$os = "centos-bento"
|
|
\$subnet = "10.0.20"
|
|
\$network_plugin = "flannel"
|
|
\$inventory = "$INV"
|
|
\$shared_folders = { 'temp/docker_rpms' => "/var/cache/yum/x86_64/7/docker-ce/packages" }
|
|
EOF
|
|
|
|
# make the rpm cache
|
|
mkdir -p temp/docker_rpms
|
|
|
|
vagrant up
|
|
|
|
# make a copy of the downloaded docker rpm, to speed up the next provisioning run
|
|
scp kub-1:/var/cache/yum/x86_64/7/docker-ce/packages/* temp/docker_rpms/
|
|
|
|
# copy kubectl access configuration in place
|
|
mkdir $HOME/.kube/ &> /dev/null
|
|
ln -s $PWD/$INV/artifacts/admin.conf $HOME/.kube/config
|
|
# make the kubectl binary available
|
|
sudo ln -s $PWD/$INV/artifacts/kubectl /usr/local/bin/kubectl
|
|
#or
|
|
export PATH=$PATH:$PWD/$INV/artifacts
|
|
```
|
|
|
|
If a vagrant run failed and you've made some changes to fix the issue causing
|
|
the fail, here is how you would re-run ansible:
|
|
|
|
```ShellSession
|
|
ansible-playbook -vvv -i .vagrant/provisioners/ansible/inventory/vagrant_ansible_inventory cluster.yml
|
|
```
|
|
|
|
If all went well, you check if it's all working as expected:
|
|
|
|
```ShellSession
|
|
kubectl get nodes
|
|
```
|
|
|
|
The output should look like this:
|
|
|
|
```ShellSession
|
|
$ kubectl get nodes
|
|
NAME STATUS ROLES AGE VERSION
|
|
kub-1 Ready control-plane,master 4m37s v1.22.5
|
|
kub-2 Ready control-plane,master 4m7s v1.22.5
|
|
kub-3 Ready <none> 3m7s v1.22.5
|
|
```
|
|
|
|
Another nice test is the following:
|
|
|
|
```ShellSession
|
|
kubectl get pods --all-namespaces -o wide
|
|
```
|
|
|
|
Which should yield something like the following:
|
|
|
|
```ShellSession
|
|
$ kubectl get pods --all-namespaces -o wide
|
|
NAMESPACE NAME READY STATUS RESTARTS AGE IP NODE NOMINATED NODE READINESS GATES
|
|
kube-system coredns-8474476ff8-m2469 1/1 Running 0 2m45s 10.233.65.2 kub-2 <none> <none>
|
|
kube-system coredns-8474476ff8-v5wzj 1/1 Running 0 2m41s 10.233.64.3 kub-1 <none> <none>
|
|
kube-system dns-autoscaler-5ffdc7f89d-76tnv 1/1 Running 0 2m43s 10.233.64.2 kub-1 <none> <none>
|
|
kube-system kube-apiserver-kub-1 1/1 Running 1 4m54s 10.0.20.101 kub-1 <none> <none>
|
|
kube-system kube-apiserver-kub-2 1/1 Running 1 4m33s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system kube-controller-manager-kub-1 1/1 Running 1 5m1s 10.0.20.101 kub-1 <none> <none>
|
|
kube-system kube-controller-manager-kub-2 1/1 Running 1 4m33s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system kube-flannel-9xgf5 1/1 Running 0 3m10s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system kube-flannel-l8jbl 1/1 Running 0 3m10s 10.0.20.101 kub-1 <none> <none>
|
|
kube-system kube-flannel-zss4t 1/1 Running 0 3m10s 10.0.20.103 kub-3 <none> <none>
|
|
kube-system kube-multus-ds-amd64-bhpc9 1/1 Running 0 3m2s 10.0.20.103 kub-3 <none> <none>
|
|
kube-system kube-multus-ds-amd64-n6vl8 1/1 Running 0 3m2s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system kube-multus-ds-amd64-qttgs 1/1 Running 0 3m2s 10.0.20.101 kub-1 <none> <none>
|
|
kube-system kube-proxy-2x4jl 1/1 Running 0 3m33s 10.0.20.101 kub-1 <none> <none>
|
|
kube-system kube-proxy-d48r7 1/1 Running 0 3m33s 10.0.20.103 kub-3 <none> <none>
|
|
kube-system kube-proxy-f45lp 1/1 Running 0 3m33s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system kube-scheduler-kub-1 1/1 Running 1 4m54s 10.0.20.101 kub-1 <none> <none>
|
|
kube-system kube-scheduler-kub-2 1/1 Running 1 4m33s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system nginx-proxy-kub-3 1/1 Running 0 3m33s 10.0.20.103 kub-3 <none> <none>
|
|
kube-system nodelocaldns-cg9tz 1/1 Running 0 2m41s 10.0.20.102 kub-2 <none> <none>
|
|
kube-system nodelocaldns-htswt 1/1 Running 0 2m41s 10.0.20.103 kub-3 <none> <none>
|
|
kube-system nodelocaldns-nsp7s 1/1 Running 0 2m41s 10.0.20.101 kub-1 <none> <none>
|
|
local-path-storage local-path-provisioner-66df45bfdd-km4zg 1/1 Running 0 2m54s 10.233.66.2 kub-3 <none> <none>
|
|
```
|