PlatformEngr® Progressive delivery course

Module 01 · Foundations · 25 min · You are: Platform team

Install Argo Rollouts from Git

In this module

  • You learn: What Argo Rollouts installs (a controller and five resource kinds), why it is a platform add-on, and how Argo CD already understands a Rollout's health.
  • You do: Add an Application for the argo-rollouts Helm chart with a values file in your platform repo, let the root app install it, and check it with the CRD list, the dashboard and the plugin.
  • You end with: Argo Rollouts v1.10.2 runs in the argo-rollouts namespace, installed from your platform repo. Nothing uses it yet: checkout is still a Deployment.

The controller is shared by every team, so it belongs to the platform team and goes in like Traefik did: an upstream chart, pinned, with values in Git. Teams then write Rollout resources in their own charts and never touch the controller.

Open this module in the interactive course to track progress and switch tracks.

Release path practised: Committed → Deployed → Exposed → Measured → Decided → Promoted → Rolled back

Concepts

Rollout
Like a Deployment (same selector and pod template), plus a strategy with steps. It owns ReplicaSets; one per version.
AnalysisTemplate
A reusable check: which metric, how often, and what counts as success. ClusterAnalysisTemplate is the cluster-wide version.
AnalysisRun
One execution of a template during one release. It ends Successful, Failed, Error or Inconclusive.
The kubectl plugin
kubectl argo rollouts shows a Rollout's steps, versions and analysis, and can promote or abort. It talks to the cluster, not to Argo CD.

Files to know

Open them in your editor or on GitHub; the bullets say what to look at.

Steps

1. Add the Application and its values

Copy the Application, then write the values file yourself. Its first version only turns on the read-only dashboard.

cd ~/work/argocd-platform && git switch -q main && git pull -q
mkdir -p argo-rollouts
cp ~/work/progressive-delivery/platform/apps/argo-rollouts.yaml apps/
perl -pi -e 's/YOUR_GITHUB_USER/YOUR_GITHUB_USER/g' apps/argo-rollouts.yaml     # your GitHub user instead of the placeholder

Create ~/work/argocd-platform/argo-rollouts/values.yaml

# Argo Rollouts Helm values: the platform team's configuration of the controller.
controller:
  replicas: 1                      # one is enough on a laptop; two in production

# The read-only dashboard: kubectl -n argo-rollouts port-forward svc/argo-rollouts-dashboard 3100:3100
dashboard:
  enabled: true
  readonly: true                   # watch here, change through Git

Run the check from the Argo CD course. Expect ok:

scripts/validate.sh
git add -A && git commit -qm "Argo Rollouts v1.10.2" && git push -q

2. Let the root app install it

The root Application sees a new file in apps/ and creates the argo-rollouts Application, which installs the chart.

argocd app get root --refresh >/dev/null

Wait for the install. Expect Synced and Healthy within two minutes:

argocd app wait argo-rollouts --sync --health --timeout 300

Two pods, 1/1 each: the controller and the dashboard.

kubectl -n argo-rollouts get pods

3. What it added to Kubernetes

Five new resource kinds, all in the argoproj.io group next to Argo CD's.

Expect rollouts, analysisruns, analysistemplates, clusteranalysistemplates and experiments, plus Argo CD's applications, applicationsets and appprojects:

kubectl get crd -o name | grep argoproj.io

4. The plugin and the dashboard

Two ways to watch releases. Neither changes anything you would not also change in Git.

The plugin reads the cluster. Expect No resources found: no Rollout exists yet.

kubectl argo rollouts list rollouts -A

In a second terminal, forward the dashboard and leave it running:

kubectl -n argo-rollouts port-forward svc/argo-rollouts-dashboard 3100:3100

Open http://localhost:3100/rollouts. Find: 1. the namespace selector at the top (pick checkout-dev), 2. the empty list (no Rollouts yet), 3. no Promote or Abort buttons anywhere: the dashboard is read-only, as the values said. You will come back here in module 02.

Simplified for the course

One controller replica. In production run two (controller.replicas: 2): they elect a leader, so a release keeps going when a node is drained.

The chart installs the CRDs and keeps them on uninstall (keepCRDs: true). Bump the chart like Argo CD: read the release notes, change targetRevision in a pull request.

Break it on purpose: A version that does not exist

A platform engineer bumps the chart "to the latest" from memory and gets the number wrong. The push goes through; CI only checks YAML.

You practise: reading why an Application cannot render, and seeing that a broken source stops changes without breaking what runs.

1. Break it

cd ~/work/argocd-platform && git switch -q main && git pull -q

Change in ~/work/argocd-platform/apps/argo-rollouts.yaml

-       targetRevision: 2.43.12            # Argo Rollouts v1.10.2
+       targetRevision: 2.43.99
git add -A && git commit -qm "Bump Argo Rollouts" && git push -q
argocd app get root --refresh >/dev/null; sleep 20; argocd app get argo-rollouts --refresh >/dev/null

2. What you see

Expected: argo-rollouts Synced on the new version. You get: argocd app get argo-rollouts shows Sync Status: Unknown and a condition. The controller pods still run the old version.

Look at the Application's conditions in the UI (the red banner on the app) or with argocd app get argo-rollouts.

3. Fix it yourself

Find what Argo CD cannot do, and put the Application back to a version that exists. Done when argo-rollouts is Synced and Healthy again.

Hint 1

The condition is a ComparisonError: Argo CD could not render the source, so it has nothing to compare with.

Hint 2

Read the message: Helm cannot find that chart version in the repository. Undo the commit.

The fix
cd ~/work/argocd-platform && git switch -q main && git pull -q
git revert --no-edit HEAD && git push -q
argocd app get root --refresh >/dev/null; sleep 20; argocd app get argo-rollouts --refresh >/dev/null
argocd app wait argo-rollouts --sync --health --timeout 300
After you fixed it: the five questions
What failed?
The source: a chart version that is not in the repository.
How was it detected?
Argo CD: Sync Status: Unknown with a ComparisonError naming the chart version. Nothing in the cluster changed.
Who responds?
The platform team, who own the Application.
How is it corrected?
A revert in the platform repo.
How is it prevented?
Let CI render the chart (helm template with the pinned version) before merge, and let a bot such as Renovate propose bumps with versions that exist.

Quiz

  1. Why is the Argo Rollouts controller in the platform repo and not in checkout-app?
    • Because only the platform team can run kubectl
    • Because it is shared by every team and installs cluster-wide resources (CRDs)
    • Because Argo CD cannot install controllers from a product repo
  2. The dashboard is read-only. How does a team promote a release?
    • Through Git only: there is no other way
    • With the kubectl plugin or an Argo CD action, as an operation on a release that Git already describes
    • By editing the Rollout in the cluster with kubectl edit
  3. A chart version in an Application does not exist. What happens to the running controller?
    • Nothing: it keeps running the last synced version
    • Argo CD deletes it because the source is empty
    • It is downgraded to the newest version that exists
Answers
  1. Because it is shared by every team and installs cluster-wide resources (CRDs) Cluster-wide, shared pieces are the platform's; teams use them through their own resources.
  2. With the kubectl plugin or an Argo CD action, as an operation on a release that Git already describes Git says which version; promote and abort are decisions about how far that version has gone.
  3. Nothing: it keeps running the last synced version Argo CD cannot render, so it changes nothing.

Field note

Teams often install Argo Rollouts with a kubectl apply of the release manifest, "just to try it", and a year later nobody knows which version runs. Installing it the same way as everything else costs five minutes once.

Report a problem with this module