Inference Public

Platform engineering

Argo CD and the record of what is running.

GitOps makes a repository the place where a deployment is requested, reviewed and recorded. That changes how releases are approved, how drift is handled and how a service is rolled back.

Rear view of two monitors and their connections at an empty workstation.

Separate building from deploying

In a GitOps arrangement the build pipeline compiles, tests and publishes an image, and then proposes a change to the configuration that describes each environment. Argo CD watches that configuration and brings the cluster into line with it. Because Argo CD pulls the change, the build pipeline needs no credentials for the cluster and no access to the Argo CD API.

For organisations with formal change control this is useful in a practical way. Fewer systems hold credentials that can reach production, and the approved pull request becomes the change record, with the reviewer, the time and the exact difference all kept in one place.

Choose the sync behaviour deliberately

Argo CD can synchronise automatically whenever the repository and the cluster differ. Two options are off by default. Pruning deletes resources that have been removed from Git, and self-healing reverts manual changes made directly to the cluster. Without self-healing, Argo CD attempts one sync for each commit, and a failed sync for the same commit is not retried. The repository is checked roughly every three minutes.

A common pattern is fully automated sync with pruning and self-healing in development, and automated sync with self-healing in production, where removals are reviewed before pruning is applied. Services with a formal approval step may keep production on manual sync. Whichever you choose, note that Argo CD will not perform a rollback while automated sync is enabled, so the rollback runbook should describe reverting the commit in Git or switching automated sync off first.

Use sync windows for change control

Sync windows allow or block synchronisation on a schedule, and they can be applied to selected applications, namespaces or clusters. Where an allow window and a deny window are both active, the deny window wins. Windows let you encode agreed change periods and freezes, such as a payment run or a statutory deadline, in configuration that is itself reviewed. Argo CD can permit manual syncs during a window, but check how that override behaves in your version before writing it into an incident procedure.

Structure repositories around ownership

Argo CD projects restrict which repositories and destinations an application may use, which gives each team a bounded space in a shared cluster. Let each application team own its configuration folder, let the platform team own cluster add-ons and policies, and use code-owner rules so that the right people review each change. Promote a release between environments by changing the image reference in a pull request, which keeps every approval visible.

Self-managed Argo CD or the EKS capability

Since late 2025 AWS has offered Argo CD as a managed EKS capability. It runs in the AWS control plane rather than on your worker nodes, integrates with IAM Identity Center so that groups map to Argo CD roles, and registers target clusters by their EKS ARN. It supports a subset of upstream settings through the standard argocd-cm ConfigMap, and it only deploys to EKS clusters.

The managed capability suits an estate that is entirely on EKS and a team that would rather not patch and scale another component. Self-managed Argo CD remains the better fit where you deploy to clusters outside EKS or depend on upstream features outside the supported subset. Compare the current pricing with the engineering time you would otherwise spend.

Hand it over with the service

The receiving team should be able to see what version is running in each environment, roll back by reverting a commit, pause automated sync during an incident, name who approves production changes and know where drift alerts go. Rehearse each of these before acceptance.

Read about platform engineering or discuss your path to production.

Further reading

Argo CD: automated sync policy

Argo CD: sync windows

Amazon EKS: continuous deployment with Argo CD

Amazon EKS: comparing the Argo CD capability with self-managed Argo CD

Related articles

What does your team still own on Amazon EKS?

A deployment pipeline that produces its own evidence.

All insights

Start a conversation

Let’s talk about what you need to build.

Bring the service, the constraint or the engineering question. We’ll work through a sensible starting point.

Discuss a project