kubespray/docs/ansible.md

8.9 KiB

Ansible variables

Inventory

The inventory is composed of 3 groups:

  • kube_node : list of kubernetes nodes where the pods will run.
  • kube_control_plane : list of servers where kubernetes control plane components (apiserver, scheduler, controller) will run.
  • etcd: list of servers to compose the etcd server. You should have at least 3 servers for failover purpose.

Note: do not modify the children of k8s_cluster, like putting the etcd group into the k8s_cluster, unless you are certain to do that and you have it fully contained in the latter:

k8s_cluster ⊂ etcd => kube_node ∩ etcd = etcd

When kube_node contains etcd, you define your etcd cluster to be as well schedulable for Kubernetes workloads. If you want it a standalone, make sure those groups do not intersect. If you want the server to act both as control-plane and node, the server must be defined on both groups kube_control_plane and kube_node. If you want a standalone and unschedulable control plane, the server must be defined only in the kube_control_plane and not kube_node.

There are also two special groups:

Below is a complete inventory example:

## Configure 'ip' variable to bind kubernetes services on a
## different ip than the default iface
node1 ansible_host=95.54.0.12 ip=10.3.0.1
node2 ansible_host=95.54.0.13 ip=10.3.0.2
node3 ansible_host=95.54.0.14 ip=10.3.0.3
node4 ansible_host=95.54.0.15 ip=10.3.0.4
node5 ansible_host=95.54.0.16 ip=10.3.0.5
node6 ansible_host=95.54.0.17 ip=10.3.0.6

[kube_control_plane]
node1
node2

[etcd]
node1
node2
node3

[kube_node]
node2
node3
node4
node5
node6

[k8s_cluster:children]
kube_node
kube_control_plane

Group vars and overriding variables precedence

The group variables to control main deployment options are located in the directory inventory/sample/group_vars. Optional variables are located in the inventory/sample/group_vars/all.yml. Mandatory variables that are common for at least one role (or a node group) can be found in the inventory/sample/group_vars/k8s_cluster.yml. There are also role vars for docker, kubernetes preinstall and control plane roles. According to the ansible docs, those cannot be overridden from the group vars. In order to override, one should use the -e runtime flags (most simple way) or other layers described in the docs.

Kubespray uses only a few layers to override things (or expect them to be overridden for roles):

Layer Comment
role defaults provides best UX to override things for Kubespray deployments
inventory vars Unused
inventory group_vars Expects users to use all.yml,k8s_cluster.yml etc. to override things
inventory host_vars Unused
playbook group_vars Unused
playbook host_vars Unused
host facts Kubespray overrides for internal roles' logic, like state flags
play vars Unused
play vars_prompt Unused
play vars_files Unused
registered vars Unused
set_facts Kubespray overrides those, for some places
role and include vars Provides bad UX to override things! Use extra vars to enforce
block vars (only for tasks in block) Kubespray overrides for internal roles' logic
task vars (only for the task) Unused for roles, but only for helper scripts
extra vars (always win precedence) override with ansible-playbook -e @foo.yml

Ansible tags

The following tags are defined in playbooks:

Tag name Used for
apps K8s apps definitions
azure Cloud-provider Azure
bastion Setup ssh config for bastion
bootstrap-os Anything related to host OS configuration
calico Network plugin Calico
canal Network plugin Canal
cloud-provider Cloud-provider related tasks
docker Configuring docker for hosts
download Fetching container images to a delegate host
etcd Configuring etcd cluster
etcd-pre-upgrade Upgrading etcd cluster
etcd-secrets Configuring etcd certs/keys
etchosts Configuring /etc/hosts entries for hosts
facts Gathering facts and misc check results
flannel Network plugin flannel
gce Cloud-provider GCP
k8s-pre-upgrade Upgrading K8s cluster
k8s-secrets Configuring K8s certs/keys
kube-apiserver Configuring static pod kube-apiserver
kube-controller-manager Configuring static pod kube-controller-manager
kubectl Installing kubectl and bash completion
kubelet Configuring kubelet service
kube-proxy Configuring static pod kube-proxy
kube-scheduler Configuring static pod kube-scheduler
localhost Special steps for the localhost (ansible runner)
master Configuring K8s master node role
netchecker Installing netchecker K8s app
network Configuring networking plugins for K8s
nginx Configuring LB for kube-apiserver instances
node Configuring K8s minion (compute) node role
openstack Cloud-provider OpenStack
preinstall Preliminary configuration steps
resolvconf Configuring /etc/resolv.conf for hosts/apps
upgrade Upgrading, f.e. container images/binaries
upload Distributing images/binaries across hosts
weave Network plugin Weave
ingress_alb AWS ALB Ingress Controller
ambassador Ambassador Ingress Controller

Note: Use the bash scripts/gen_tags.sh command to generate a list of all tags found in the codebase. New tags will be listed with the empty "Used for" field.

Example commands

Example command to filter and apply only DNS configuration tasks and skip everything else related to host OS configuration and downloading images of containers:

ansible-playbook -i inventory/sample/hosts.ini cluster.yml --tags preinstall,facts --skip-tags=download,bootstrap-os

And this play only removes the K8s cluster DNS resolver IP from hosts' /etc/resolv.conf files:

ansible-playbook -i inventory/sample/hosts.ini -e dns_mode='none' cluster.yml --tags resolvconf

And this prepares all container images locally (at the ansible runner node) without installing or upgrading related stuff or trying to upload container to K8s cluster nodes:

ansible-playbook -i inventory/sample/hosts.ini cluster.yml \
    -e download_run_once=true -e download_localhost=true \
    --tags download --skip-tags upload,upgrade

Note: use --tags and --skip-tags wise and only if you're 100% sure what you're doing.

Bastion host

If you prefer to not make your nodes publicly accessible (nodes with private IPs only), you can use a so called bastion host to connect to your nodes. To specify and use a bastion, simply add a line to your inventory, where you have to replace x.x.x.x with the public IP of the bastion host.

[bastion]
bastion ansible_host=x.x.x.x

For more information about Ansible and bastion hosts, read Running Ansible Through an SSH Bastion Host

Mitogen

You can use mitogen to speed up kubespray.

Beyond ansible 2.9

Ansible project has decided, in order to ease their maintenance burden, to split between two projects which are now joined under the Ansible umbrella.

Ansible-base (2.10.x branch) will contain just the ansible language implementation while ansible modules that were previously bundled into a single repository will be part of the ansible 3.x package. Pleasee see this blog post that explains in detail the need and the evolution plan.

Note: this change means that ansible virtual envs cannot be upgraded with pip install -U. You first need to uninstall your old ansible (pre 2.10) version and install the new one.

pip uninstall ansible
cd kubespray/
pip install -U .

Note: some changes needed to support ansible 2.10+ are not backwards compatible with 2.9 Kubespray needs to evolve and keep pace with upstream ansible and will be forced to eventually drop 2.9 support. Kubespray CIs use only the ansible version specified in the requirements.txt and while the ansible_version.yml may allow older versions to be used, these are not exercised in the CI and compatibility is not guaranteed.