Immutable Kubernetes configuration with Ansible

28 Jan 2019

This post touches on a key component of my Managing Kubernetes is Easy With Ansible talk that I gave at AnsibleFest 2018. Since giving that talk, I’ve also solved some of the unforeseen consequences, and go into further detail here.

There are a number of problems associated with managing configuration (in the form of configmaps and secrets) through the default mechanism.

We can solve these problems through immutable configmaps and secrets that are named based on their contents, such that if the contents change, their name changes.

The kubectl tool has part of a solution for that, in that you can pass --append-hash when creating a configmap or secret, but only when using kubectl create, it’s not useful when applying resource definitions from files. However, this idea was the inspiration to add an append_hash parameter to Ansible’s k8s module, and this is part of the solution.

Another part of the solution is the k8s_config_resource_name filter plugin, which takes a configmap definition and returns the full name with the hash added.

At this point we can define dicts of configmaps and secrets, and also refer to the full resource name in deployments or similar resources.

In inventory we might have something like:

  env: "{{ lookup('template', kube_resource_template_dir + 'env-configmap.yml') | from_yaml }}"
  env: "{{ lookup('template', kube_resource_template_dir + 'env-secrets.yml') | from_yaml }}"

which then gets referenced in the deployment with:

apiVersion: apps/v1
kind: Deployment
  name: {{ kube_resource_name }}
  namespace: {{ kube_resource_namespace }}
        - envFrom:
            - configMapRef:
                name: {{ kube_resource_configmaps.env | k8s_config_resource_name }}
            - secretRef:
                name: {{ kube_resource_secrets.env | k8s_config_resource_name }}

Its best to create the secrets and configmaps first, as the pods created by the deployment will depend on them being present to be able to start - then you can use the wait functionality that the k8s module gains with Ansible 2.8 to check the pods start correctly.

At this point, we have met our original goals - changing a configmap or secret will change its name, and so the deployment will change to point to the new name, so new pods will be created using the new configuration. The replicaset to which kubectl rollback deploy will roll back will contain the old configuration names, and so rolling back a deployment will also roll back the configuration.

However, my presentation ignores two issues (the first I was aware of but hadn’t yet solved, the second my colleague highlighted as an obstacle to moving to the new way of working)

The solution to both is simple to describe, and harder to implement with the current tools, but possible.

Kubernetes has the concept of resource owners: if a resource that owns another resource is deleted, the owned resource is also deleted. So pods are owned by replicasets which are in turn owned by deployments - deleting the deployment removes all replicasets associated with the deployment, and all pods associated with those replicasets.

We can use this to add owner references to configmaps and secrets - we find the replicaset associated with the recent deployment, and set that resource as the owner in the configmap/secret. This means that when replicasets are retired (Kubernetes keeps ten replicasets by default, but this is configurable with spec.revisionHistoryLimit) the associated configmaps and secrets are also retired.

For diffs, first we lookup the annotation of the previous deployment and the current deployment. We then have to find all replicasets (as there’s no way to search by annotation, unfortunately) and then select the previous replicaset and current replicaset using the same annotation.

Adding a new label, such as kube_resource_prefix, to our configmaps allows us to iterate over our kube_resource_configmaps dict (the label_selectors argument to k8s_facts is useful here), each time looking for all configmaps with the label of the current configmap, and then finding the configmap with the owner reference set to the previous replicaset and the configmap owned by the current replicaset. Exactly the same technique holds for secrets too. Once we’ve found the before and after, we can display the differences using the debug module (an explicit diff module might be useful).

I’ve updated the ansible-role-kube-resource role with this new functionality so that it’s easier to use, it does rely on all resources having kube_resource_name and kube_resource_namespace correctly set. All of the owner reference based functionality relies on replicasets updating when deployments change - so isn’t currently useful for e.g. statefulsets.

In other news on that role, I’ve now added a molecule test suite, inspired by Jeff Geerling (@geerlingguy)’s talk at AnsibleFest and his super helpful molecule blog post. It relies on an existing kubernetes platform being set up and configured, but given that Docker comes with Kubernetes (on some platforms at least), and that minikube exists, that shouldn’t be insurmountable.

There are quite a few features from Ansible 2.8 that are used in the role but have been backported into the role - kubernetes resource validation, waiting for resources to meet the desired state, append_hash and k8s_config_resource_name for example. Even if you choose not to use the role, you can make use of the same techniques through setting and using module_utils and filter_plugins configuration directives in ansible.cfg.