GitOps using FluxCD (v2)

FluxCD is a GitOps operator for Kubernetes. It synchronizes the status of the cluster from manifests allocated in different repositories (Git or Helm). This approach fits perfectly with External Secrets on clusters which are dynamically created, to get credentials with no manual intervention from the beginning.

Advantages

This approach has several advantages as follows:

Homogenize environments allowing developers to use the same toolset in Kind in the same way they do in the cloud provider distributions such as EKS or GKE. This accelerates the development
Reduce security risks, because credentials can be easily obtained, so temptation to store them locally is reduced.
Application compatibility increase: Applications are deployed in different ways, and sometimes they need to share credentials. This can be done using External Secrets as a wire for them at real time.
Automation by default oh, come on!

The approach

FluxCD is composed by several controllers dedicated to manage different custom resources. The most important ones are Kustomization (to clarify, Flux one, not Kubernetes' one) and HelmRelease to deploy using the approaches of the same names.

External Secrets can be deployed using Helm as explained here. The deployment includes the CRDs if enabled on the values.yaml, but after this, you need to deploy some SecretStore to start getting credentials from your secrets manager with External Secrets.

The idea of this guide is to deploy the whole stack, using flux, needed by developers not to worry about the credentials, but only about the application and its code.

The problem

This can sound easy, but External Secrets is deployed using Helm, which is managed by the HelmController, and your custom resources, for example a ClusterSecretStore and the related Secret, are often deployed using a kustomization.yaml, which is deployed by the KustomizeController.

Both controllers manage the resources independently, at different moments, with no possibility to wait each other. This means that we have a wonderful race condition where sometimes the CRs (SecretStore,ClusterSecretStore...) tries to be deployed before than the CRDs needed to recognize them.

A second, subtler race exists around the admission webhook. External Secrets ships a ValidatingWebhookConfiguration that is registered in the API server as soon as the HelmRelease is applied, before the webhook pod has had time to start serving. If a Kustomization tries to apply an ExternalSecret or ClusterSecretStore in that brief window, the API server performs a dry-run validation, the webhook endpoint returns connection refused, and the reconciliation fails with:

Internal error occurred: failed calling webhook "validate.externalsecret.external-secrets.io": ...
dial tcp <ip>:443: connect: connection refused

Flux retries after interval (typically 10 minutes), so everything works on the second attempt, but the initial deployment always fails. The fix is a three-level dependency chain that ensures the webhook pod is healthy before any CR is applied.

The solution

Let's see the conditions to start working on a solution:

The External Secrets operator is deployed with Helm, and admits disabling the CRDs deployment
The race condition only affects the deployment of CustomResourceDefinition and the CRs needed later
CRDs can be deployed directly from the Git repository of the project using a Flux Kustomization
The operator HelmRelease is wrapped in its own Flux Kustomization with wait: true so it only reports Ready after all deployed resources (including the webhook pod) are healthy
Required CRs can be deployed using a Flux Kustomization that depends on the operator Kustomization, not just the CRDs, guaranteeing the webhook is serving before any CR dry-run is attempted
All previous manifests can be applied with a Kubernetes kustomization

The dependency chain is:

external-secrets-crds --> external-secrets-operator (wait: true) --> external-secrets-crs

Create the main kustomization

To have a better view of things needed later, the first manifest to be created is the kustomization.yaml

{% include 'gitops/kustomization.yaml' %}

Create the secret

To access your secret manager, External Secrets needs some credentials. They are stored inside a Secret, which is intended to be deployed by automation as a good practise. This time, a placeholder called secret-token.yaml is show as an example:

# The namespace.yaml first
{% include 'gitops/namespace.yaml' %}

{% include 'gitops/secret-token.yaml' %}

Creating the references to repositories

Create a manifest called repositories.yaml to store the references to external repositories for Flux

{% include 'gitops/repositories.yaml' %}

Deploy the CRDs

As mentioned, CRDs can be deployed using the official Helm package, but to solve the race condition, they will be deployed from our git repository using a Kustomization manifest called deployment-crds.yaml as follows:

{% include 'gitops/deployment-crds.yaml' %}

Note the wait: true field. This ensures the Kustomization only reports Ready after all CRDs have been fully established in the API server, so the operator can register its validation webhooks and controllers cleanly.

Deploy the operator

The operator HelmRelease is placed inside an operator/ subdirectory and wrapped with its own Flux Kustomization. Create a manifest called deployment-operator.yaml:

{% include 'gitops/deployment-operator.yaml' %}

The wait: true here is the key to solving the webhook race condition: Flux will not mark external-secrets-operator as Ready until every resource the HelmRelease created -- including the webhook Deployment -- has reached a healthy state. Only then will the CRs Kustomization be allowed to proceed.

Inside the operator/ subdirectory, place the HelmRelease manifest (operator/deployment.yaml):

{% include 'gitops/operator/deployment.yaml' %}

Deploy the CRs

Now, be ready for the arcane magic. Create a Kustomization manifest called deployment-crs.yaml with the following content:

{% include 'gitops/deployment-crs.yaml' %}

There are several interesting details to see here, that finally solves the race condition:

The dependsOn field now points to external-secrets-operator rather than external-secrets-crds. This dependency forces this deployment to wait for the operator (including its webhook) to be fully ready before any CR is applied, eliminating the webhook race condition.
retryInterval: 1m makes Flux retry quickly if the very first reconcile still catches a brief startup window.
The reference to the place where to find the CRs
```
path: ./infrastructure/external-secrets/crs
sourceRef:
kind: GitRepository
name: flux-system
```
Custom Resources will be searched in the relative path ./infrastructure/external-secrets/crs of the GitRepository called flux-system, which is a reference to the same repository that FluxCD watches to synchronize the cluster. With fewer words, a reference to itself, but going to another directory called crs

Of course, allocate inside the mentioned path ./infrastructure/external-secrets/crs, all the desired CRs to be deployed, for example, a manifest clusterSecretStore.yaml to reach your Hashicorp Vault as follows:

{% include 'gitops/crs/clusterSecretStore.yaml' %}

Results

At the end, the required files tree is:

./infrastructure/external-secrets/
  kustomization.yaml
  namespace.yaml
  secret-token.yaml
  repositories.yaml
  deployment-crds.yaml
  deployment-operator.yaml
  operator/
    kustomization.yaml
    deployment.yaml
  deployment-crs.yaml
  crs/
    kustomization.yaml
    clusterSecretStore.yaml

gitops-using-fluxcd.md 7.8 KB Permalink Histórico Raw