Contributors to this page
Last update:

Extensibility overview

Initially, everything was developed in-tree in the Gardener project. All cloud providers and the configuration for all the supported operating systems were released together with the Gardener core itself. But as the project grew, it got more and more difficult to add new providers and maintain the existing code base. As a consequence and in order to become agile and flexible again, we proposed GEP-1 (Gardener Enhancement Proposal). The document describes an out-of-tree extension architecture that keeps the Gardener core logic independent of provider-specific knowledge (similar to what Kubernetes has achieved with out-of-tree cloud providers or with CSI volume plugins).

Basic concepts

Gardener keeps running in the “garden cluster” and implements the core logic of shoot cluster reconciliation/deletion. Extensions are Kubernetes controllers themselves (like Gardener) and run in the seed clusters. As usual, we try to use Kubernetes wherever applicable. We rely on Kubernetes extension concepts in order to enable extensibility for Gardener. The main ideas of GEP-1 are the following:

  1. During the shoot reconciliation process Gardener will write CRDs into the seed cluster that are watched and managed by the extension controllers. They will reconcile (based on the .spec) and report whether everything went well or errors occurred in the CRD’s .status field.

  2. Gardener keeps deploying the provider-independent control plane components (etcd, kube-apiserver, etc.). However, some of these components might still need little customization by providers, e.g., additional configuration, flags, etc. In this case, the extension controllers register webhooks in order to manipulate the manifests.

Example 1:

Gardener creates a new AWS shoot cluster and requires the preparation of infrastructure in order to proceed (networks, security groups, etc.). It writes the following CRD into the seed cluster:

apiVersion: extensions.gardener.cloud/v1alpha1
kind: Infrastructure
metadata:
  name: infrastructure
  namespace: shoot--core--aws-01
spec:
  type: aws
  providerConfig:
    apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
    kind: InfrastructureConfig
    networks:
      vpc:
        cidr: 10.250.0.0/16
      internal:
      - 10.250.112.0/22
      public:
      - 10.250.96.0/22
      workers:
      - 10.250.0.0/19
    zones:
    - eu-west-1a
  dns:
    apiserver: api.aws-01.core.example.com
  region: eu-west-1
  secretRef:
    name: my-aws-credentials
  sshPublicKey: |
    base64(key)

Please note that the .spec.providerConfig is a raw blob and not evaluated or known in any way by Gardener. Instead, it was specified by the user (in the Shoot resource) and just “forwarded” to the extension controller. Only the AWS controller understands this configuration and will now start provisioning/reconciling the infrastructure. It reports in the .status field the result:

status:
  observedGeneration: ...
  state: ...
  lastError: ..
  lastOperation: ...
  providerStatus:
    apiVersion: aws.provider.extensions.gardener.cloud/v1alpha1
    kind: InfrastructureStatus
    vpc:
      id: vpc-1234
      subnets:
      - id: subnet-acbd1234
        name: workers
        zone: eu-west-1
      securityGroups:
      - id: sg-xyz12345
        name: workers
    iam:
      nodesRoleARN: <some-arn>
      instanceProfileName: foo
    ec2:
      keyName: bar

Gardener waits until the .status.lastOperation/.status.lastError indicates that the operation reached a final state and either continuous with the next step or stops and reports the potential error. The extension-specific output in .status.providerStatus is - similar to .spec.providerConfig - not evaluated and simply forwarded to CRDs in subsequent steps.

Example 2:

Gardener deploys the control plane components into the seed cluster, e.g. the kube-controller-manager deployment with the following flags:

apiVersion: apps/v1
kind: Deployment
...
spec:
  template:
    spec:
      containers:
      - command:
        - /usr/local/bin/kube-controller-manager
        - --allocate-node-cidrs=true
        - --attach-detach-reconcile-sync-period=1m0s
        - --controllers=*,bootstrapsigner,tokencleaner
        - --cluster-cidr=100.96.0.0/11
        - --cluster-name=shoot--core--aws-01
        - --cluster-signing-cert-file=/srv/kubernetes/ca/ca.crt
        - --cluster-signing-key-file=/srv/kubernetes/ca/ca.key
        - --concurrent-deployment-syncs=10
        - --concurrent-replicaset-syncs=10
...

The AWS controller requires some additional flags in order to make the cluster functional. It needs to provide a Kubernetes cloud-config and also some cloud-specific flags. Consequently, it registers a MutatingWebhookConfiguration on Deployments and adds these flags to the container:

        - --cloud-provider=external
        - --external-cloud-volume-plugin=aws
        - --cloud-config=/etc/kubernetes/cloudprovider/cloudprovider.conf

Of course, it would have needed to create a ConfigMap containing the cloud config and to add the proper volume and volumeMounts to the manifest as well.

(Please note for this special example: The Kubernetes community is also working on making the kube-controller-manager provider-independent. However, there will most probably be still components other than the kube-controller-manager which need to be adapted by extensions.)

If you are interested in writing an extension, or generally in digging deeper to find out the nitty-gritty details of the extension concepts please read GEP-1. We are truly looking forward to your feedback!

Current status

Meanwhile, the out-of-tree extension architecture of Gardener is in place and has been productively validated. We are tracking all internal and external extensions of Gardener in the repo: Gardener Extensions Library.