Skip to content

Latest commit

 

History

History
154 lines (133 loc) · 5.51 KB

README.md

File metadata and controls

154 lines (133 loc) · 5.51 KB
Raw

Design

Kraan is a kubernetes controller for processing "AddonLayer" custom resources. An AddonLayer has the following features.

  • Represents a collection of add-ons, managed as HelmRelease custom resources.
  • AddonLayer is a cluster scoped custom resource (non-namespaced).
  • Certify layers on specific k8s version. i.e All the add-ons in a layer is guaranteed to work without issues on a specified k8s version. Kraan will attempt to deploy a layer only when the cluster is at a specified version.
  • Establish dependency between layers.
  • Pause/Resume layer processing.

Components

Kraan is a combination of 3 tools. The installation of kraan will involve the installation of all these 3 tools.

layers

kraan controller is the main controller that is responsible to watching AddonLayer custom resources and bringing to its desired state. In attempting to do so, it relies on helm-controller to deploy the addon helm charts that are part of the layer. Helm operator watches and reconciles HelmRelease custom resource. Each HelmRelease custom resources represents a single addon. For more details on HelmRelease, check here. source-controller is one of the components in gitops-toolkit which helps abstracting away git interaction from kraan. By design, an AddonLayer will point to a directory in a git repository which contains the list of HelmReleases. Whenever kraans' reconciliation logic is kicked off, in order to fetch the list of helm releases that are part of that addon layer, it will reach out to source-controller api to fetch repo files instead of reaching out to git directly.

Note: In future, kraan will support packaging addons in formats other than helm charts e.g kustomization

Spec

The addons are first packaged as HelmRelease custom resources which represents a helm chart. A sample helm release custom resource is shown below. For more details on HelmRelease, check here.

# podinfo.yaml
---
apiVersion: helm.toolkit.fluxcd.io/v2beta2
kind: HelmRelease
metadata:
  name: podinfo
  namespace: default
spec:
  install:
    remediation:
      retries: -1
  upgrade:
    remediation:
      retries: -1
  chart:
    spec:
      chart: podinfo
      sourceRef:
        kind: HelmRepository
        name: podinfo
        namespace: gotk-system
      version: '>4.0.0'
  values:
    podinfo:
      service:
        enabled: true
        type: ClusterIP
      replicaCount: 1
      message: podinfo
  interval: 1m0s

Example

A sample addon layer custom resource is shown below.

apiVersion: kraan.io/v1alpha1
kind: AddonsLayer
metadata:
  name: bootstrap

spec:
  version: 1.0.0
  hold: true
  interval: 1m
  source:
    name: addons
    namespace: gotk-system
    path: ./addons/bootstrap
  prereqs:
      k8sVersion: "1.16"
---
apiVersion: kraan.io/v1alpha1
kind: AddonsLayer
metadata:
  name: base

spec:
  version: 1.0.1
  interval: 1m
  source: 
    name: addons
    namespace: gotk-system
    path: ./addons/base
  prereqs:
      k8sVersion: "1.16"
      dependsOn:
        - bootstrap@1.0.0

Reconciliation Logic

Kraan controller is built using the kubebuilder project. From a controllers' perspective, reconciliation is where it attempts to bring the current state of the layer into its desired state. The reconciliation logic is kicked off whenever there is a change detected in the AddonLayer custom resource. It could be an addition, deletion or an update to the resource. Whenever an event is received for an AddonLayer,

  • Check whether the version of the Kubernetes API deployed on the cluster is greater than or equal to prereqs.K8sVersion. If the cluster is not yet in the expected version, return back without throwing an error. Controllers' periodic sync will trigger the reconcile next time.

  • execute a custom prune logic, where any addon that is present in the cluster but NOT in the repo will be removed from the cluster. Basically any addon that is not part of the current layer spec is removed from the cluster. Note that removal of an addon does not check for dependencies between layers. If the prune is not successful, then return an error, which should trigger an exponential back off.

  • Check if the dependent layers are succesfully deployed. This is a multi step process.

    • The dependent layer is identified by a combination of name and a version. i.e base@0.1.2 means that a layer with name "base" and spec.version = 0.1.2 is expected.
    • If the dependent layer isn't found, return nil and requeue the request to process after the configured interrval.
    • If the layer is found, check whether the status of the layer is set to successfully deployed by inspecting the status subresource of that custom resource.
    • If the layer is not successfully deployed yet, return to requeue the request after the configured interrval.

  • If the dependent layer is successfully deployed, then proceed to deploy this layers' addons. Return error on failure which will trigger an exponential backoff.

At every stage, the status subresource on the layer custom resource will be updated with the latest status.