Skip to content

Latest commit

 

History

History
518 lines (412 loc) · 25.1 KB

resource-manager.md

File metadata and controls

518 lines (412 loc) · 25.1 KB
Raw

Gardener Resource Manager

Initially, the gardener-resource-manager was a project similar to the kube-addon-manager. It manages Kubernetes resources in a target cluster which means that it creates, updates, and deletes them. Also, it makes sure that manual modifications to these resources are reconciled back to the desired state.

In the Gardener project we were using the kube-addon-manager since more than two years. While we have progressed with our extensibility story (moving cloud providers out-of-tree) we had decided that the kube-addon-manager is no longer suitable for this use-case. The problem with it is that it needs to have its managed resources on its file system. This requires storing the resources in ConfigMaps or Secrets and mounting them to the kube-addon-manager pod during deployment time. The gardener-resource-manager uses CustomResourceDefinitions which allows to dynamically add, change, and remove resources with immediate action and without the need to reconfigure the volume mounts/restarting the pod.

Meanwhile, the gardener-resource-manager has evolved to a more generic component comprising several controllers and webhook handlers. It is deployed by gardenlet once per seed (in the garden namespace) and once per shoot (in the respective shoot namespaces in the seed).

Controllers

ManagedResource controller

This controller watches custom objects called ManagedResources in the resources.gardener.cloud/v1alpha1 API group. These objects contain references to secrets which itself contain the resources to be managed. The reason why a Secret is used to store the resources is that they could contain confidential information like credentials.

---
apiVersion: v1
kind: Secret
metadata:
  name: managedresource-example1
  namespace: default
type: Opaque
data:
  objects.yaml: YXBpVmVyc2lvbjogdjEKa2luZDogQ29uZmlnTWFwCm1ldGFkYXRhOgogIG5hbWU6IHRlc3QtMTIzNAogIG5hbWVzcGFjZTogZGVmYXVsdAotLS0KYXBpVmVyc2lvbjogdjEKa2luZDogQ29uZmlnTWFwCm1ldGFkYXRhOgogIG5hbWU6IHRlc3QtNTY3OAogIG5hbWVzcGFjZTogZGVmYXVsdAo=
    # apiVersion: v1
    # kind: ConfigMap
    # metadata:
    #   name: test-1234
    #   namespace: default
    # ---
    # apiVersion: v1
    # kind: ConfigMap
    # metadata:
    #   name: test-5678
    #   namespace: default
---
apiVersion: resources.gardener.cloud/v1alpha1
kind: ManagedResource
metadata:
  name: example
  namespace: default
spec:
  secretRefs:
  - name: managedresource-example1

In the above example, the controller creates two ConfigMaps in the default namespace. When a user is manually modifying them they will be reconciled back to the desired state stored in the managedresource-example secret.

It is also possible to inject labels into all the resources:

---
apiVersion: v1
kind: Secret
metadata:
  name: managedresource-example2
  namespace: default
type: Opaque
data:
  other-objects.yaml: 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
    # apiVersion: apps/v1
    # kind: Deployment
    # metadata:
    #   name: nginx-deployment
    # spec:
    #   selector:
    #     matchLabels:
    #       app: nginx
    #   replicas: 2 # tells deployment to run 2 pods matching the template
    #   template:
    #     metadata:
    #       labels:
    #         app: nginx
    #     spec:
    #       containers:
    #       - name: nginx
    #         image: nginx:1.7.9
    #         ports:
    #         - containerPort: 80

---
apiVersion: resources.gardener.cloud/v1alpha1
kind: ManagedResource
metadata:
  name: example
  namespace: default
spec:
  secretRefs:
  - name: managedresource-example2
  injectLabels:
    foo: bar

In this example the label foo=bar will be injected into the Deployment as well as into all created ReplicaSets and Pods.

Preventing Reconciliations

If a ManagedResource is annotated with resources.gardener.cloud/ignore=true then it will be skipped entirely by the controller (no reconciliations or deletions of managed resources at all). However, when the ManagedResource itself is deleted (for example when a shoot is deleted) then the annotation is not respected and all resources will be deleted as usual. This feature can be helpful to temporarily patch/change resources managed as part of such ManagedResource. Condition checks will be skipped for such ManagedResources.

Modes

The gardener-resource-manager can manage a resource in different modes. The supported modes are:

  • Ignore
    • The corresponding resource is removed from the ManagedResource status (.status.resources). No action is performed on the cluster - the resource is no longer "managed" (updated or deleted).
    • The primary use case is a migration of a resource from one ManagedResource to another one.

The mode for a resource can be specified with the resources.gardener.cloud/mode annotation. The annotation should be specified in the encoded resource manifest in the Secret that is referenced by the ManagedResource.

Resource Class

By default, gardener-resource-manager controller watches for ManagedResources in all namespaces. --namespace flag can be specified to gardener-resource-manager binary to restrict the watch to ManagedResources in a single namespace. A ManagedResource has an optional .spec.class field that allows to indicate that it belongs to given class of resources. --resource-class flag can be specified to gardener-resource-manager binary to restrict the watch to ManagedResources with the given .spec.class. A default class is assumed if no class is specified.

Conditions

A ManagedResource has a ManagedResourceStatus, which has an array of Conditions. Conditions currently include:

Condition Description
ResourcesApplied True if all resources are applied to the target cluster
ResourcesHealthy True if all resources are present and healthy
ResourcesProgressing False if all resources have been fully rolled out

ResourcesApplied may be False when:

  • the resource apiVersion is not known to the target cluster
  • the resource spec is invalid (for example the label value does not match the required regex for it)
  • ...

ResourcesHealthy may be False when:

  • the resource is not found
  • the resource is a Deployment and the Deployment does not have the minimum availability.
  • ...

ResourcesProgressing may be True when:

  • a Deployment, StatefulSet or DaemonSet has not been fully rolled out yet, i.e. not all replicas have been updated with the latest changes to spec.template.

Each Kubernetes resources has different notion for being healthy. For example, a Deployment is considered healthy if the controller observed its current revision and if the number of updated replicas is equal to the number of replicas.

The following status.conditions section describes a healthy ManagedResource:

conditions:
- lastTransitionTime: "2022-05-03T10:55:39Z"
  lastUpdateTime: "2022-05-03T10:55:39Z"
  message: All resources are healthy.
  reason: ResourcesHealthy
  status: "True"
  type: ResourcesHealthy
- lastTransitionTime: "2022-05-03T10:55:36Z"
  lastUpdateTime: "2022-05-03T10:55:36Z"
  message: All resources have been fully rolled out.
  reason: ResourcesRolledOut
  status: "False"
  type: ResourcesProgressing
- lastTransitionTime: "2022-05-03T10:55:18Z"
  lastUpdateTime: "2022-05-03T10:55:18Z"
  message: All resources are applied.
  reason: ApplySucceeded
  status: "True"
  type: ResourcesApplied

Ignoring Updates

In some cases it is not desirable to update or re-apply some of the cluster components (for example, if customization is required or needs to be applied by the end-user). For these resources, the annotation "resources.gardener.cloud/ignore" needs to be set to "true" or a truthy value (Truthy values are "1", "t", "T", "true", "TRUE", "True") in the corresponding managed resource secrets, this can be done from the components that create the managed resource secrets, for example Gardener extensions or Gardener. Once this is done, the resource will be initially created and later ignored during reconciliation.

Preserving replicas or resources in Workload Resources

The objects which are part of the ManagedResource can be annotated with

  • resources.gardener.cloud/preserve-replicas=true in case the .spec.replicas field of workload resources like Deployments, StatefulSets, etc. shall be preserved during updates.
  • resources.gardener.cloud/preserve-resources=true in case the .spec.containers[*].resources fields of all containers of workload resources like Deployments, StatefulSets, etc. shall be preserved during updates.

This can be useful if there are non-standard horizontal/vertical auto-scaling mechanisms in place. Standard mechanisms like HorizontalPodAutoscaler or VerticalPodAutoscaler will be auto-recognized by gardener-resource-manager, i.e., in such cases the annotations are not needed.

Origin

All the objects managed by the resource manager get a dedicated annotation resources.gardener.cloud/origin describing the ManagedResource object that describes this object.

By default this is in this format <namespace>/<objectname>. In multi-cluster scenarios (the ManagedResource objects are maintained in a cluster different from the one the described objects are managed), it might be useful to include the cluster identity, as well.

This can be enforced by setting the --cluster-id option. Here, several possibilities are supported:

  • given a direct value: use this as id for the source cluster
  • <cluster>: read the cluster identity from a cluster-identity config map in the kube-system namespace (attribute cluster-identity). This is automatically maintained in all clusters managed or involved in a gardener landscape.
  • <default>: try to read the cluster identity from the config map. If not found, no identity is used
  • empty string: no cluster identity is used (completely cluster local scenarios)

The format of the origin annotation with a cluster id is <cluster id>:<namespace>/<objectname>.

The default for the cluster id is the empty value (do not use cluster id).

Garbage Collector For Immutable ConfigMaps/Secrets

In Kubernetes, workload resources (e.g., Pods) can mount ConfigMaps or Secrets or reference them via environment variables in containers. Typically, when the content of such ConfigMap/Secret gets changed then the respective workload is usually not dynamically reloading the configuration, i.e., a restart is required. The most commonly used approach is probably having so-called checksum annotations in the pod template which makes Kubernetes to recreate the pod if the checksum changes. However, it has the downside that old, still running versions of the workload might not be able to properly work with the already updated content in the ConfigMap/Secret, potentially causing application outages.

In order to protect users from such outages (and to also improve the performance of the cluster), the Kubernetes community provides the "immutable ConfigMaps/Secrets feature". Enabling immutability requires ConfigMaps/Secrets to have unique names. Having unique names requires the client to delete ConfigMaps/Secret`s no longer in use.

In order to provide a similarly lightweight experience for clients (compared to the well-established checksum annotation approach), the Gardener Resource Manager features an optional garbage collector controller (disabled by default). The purpose of this controller is cleaning up such immutable ConfigMaps/Secrets if they are no longer in use.

How does the garbage collector work?

The following algorithm is implemented in the GC controller:

  1. List all ConfigMaps and Secrets labeled with resources.gardener.cloud/garbage-collectable-reference=true.
  2. List all Deployments, StatefulSets, DaemonSets, Jobs, CronJobs, Pods and for each of them
    1. iterate over the .metadata.annotations and for each of them
      1. If the annotation key follows the reference.resources.gardener.cloud/{configmap,secret}-<hash> scheme and the value equals <name> then consider it as "in-use".
  3. Delete all ConfigMaps and Secrets not considered as "in-use".

Consequently, clients need to

  1. Create immutable ConfigMaps/Secrets with unique names (e.g., a checksum suffix based on the .data).

  2. Label such ConfigMaps/Secrets with resources.gardener.cloud/garbage-collectable-reference=true.

  3. Annotate their workload resources with reference.resources.gardener.cloud/{configmap,secret}-<hash>=<name> for all ConfigMaps/Secrets used by the containers of the respective Pods.

    ⚠️ Add such annotations to .metadata.annotations as well as to all templates of other resources (e.g., .spec.template.metadata.annotations in Deployments or .spec.jobTemplate.metadata.annotations and .spec.jobTemplate.spec.template.metadata.annotations for CronJobs. This ensures that the GC controller does not unintentionally consider ConfigMaps/Secrets as "not in use" just because there isn't a Pod referencing them anymore (e.g., they could still be used by a Deployment scaled down to 0).

ℹ️ For the last step, there is a helper function InjectAnnotations in the pkg/controller/garbagecollector/references which you can use for your convenience.

Example:

---
apiVersion: v1
kind: ConfigMap
metadata:
  name: test-1234
  namespace: default
  labels:
    resources.gardener.cloud/garbage-collectable-reference: "true"
---
apiVersion: v1
kind: ConfigMap
metadata:
  name: test-5678
  namespace: default
  labels:
    resources.gardener.cloud/garbage-collectable-reference: "true"
---
apiVersion: v1
kind: Pod
metadata:
  name: example
  namespace: default
  annotations:
    reference.resources.gardener.cloud/configmap-82a3537f: test-5678
spec:
  containers:
  - name: nginx
    image: nginx:1.14.2
    terminationGracePeriodSeconds: 2

The GC controller would delete the ConfigMap/test-1234 because it is considered as not "in-use".

ℹ️ If the GC controller is activated then the ManagedResource controller will no longer delete ConfigMaps/Secrets having the above label.

How to activate the garbage collector?

The GC controller can be activated by providing the --garbage-collector-sync-period flag with a value larger than 0 (e.g., 1h) to the Gardener Resource Manager.

TokenInvalidator

The Kubernetes community is slowly transitioning from static ServiceAccount token Secrets to ServiceAccount Token Volume Projection. Typically, when you create a ServiceAccount

apiVersion: v1
kind: ServiceAccount
metadata:
  name: default

then the serviceaccount-token controller (part of kube-controller-manager) auto-generates a Secret with a static token:

apiVersion: v1
kind: Secret
metadata:
   annotations:
      kubernetes.io/service-account.name: default
      kubernetes.io/service-account.uid: 86e98645-2e05-11e9-863a-b2d4d086dd5a)
   name: default-token-ntxs9
type: kubernetes.io/service-account-token
data:
   ca.crt: base64(cluster-ca-cert)
   namespace: base64(namespace)
   token: base64(static-jwt-token)

Unfortunately, when using ServiceAccount Token Volume Projection in a Pod, this static token is actually not used at all:

apiVersion: v1
kind: Pod
metadata:
  name: nginx
spec:
  serviceAccountName: default
  containers:
  - image: nginx
    name: nginx
    volumeMounts:
    - mountPath: /var/run/secrets/tokens
      name: token
  volumes:
  - name: token
    projected:
      sources:
      - serviceAccountToken:
          path: token
          expirationSeconds: 7200

While the Pod is now using an expiring and auto-rotated token, the static token is still generated and valid.

As of Kubernetes v1.22, there is neither a way of preventing kube-controller-manager to generate such static tokens, nor a way to proactively remove or invalidate them:

Disabling the serviceaccount-token controller is an option, however, especially in the Gardener context it may either break end-users or it may not even be possible to control such settings. Also, even if a future Kubernetes version supports native configuration of above behaviour, Gardener still supports older versions which won't get such features but need a solution as well.

This is where the TokenInvalidator comes into play: Since it is not possible to prevent kube-controller-manager from generating static ServiceAccount Secrets, the TokenInvalidator is - as its name suggests - just invalidating these tokens. It considers all such Secrets belonging to ServiceAccounts with .automountServiceAccountToken=false. By default, all namespaces in the target cluster are watched, however, this can be configured by specifying the --target-namespace flag.

apiVersion: v1
kind: ServiceAccount
metadata:
  name: my-serviceaccount
automountServiceAccountToken: false

This will result in a static ServiceAccount token secret whose token value is invalid:

apiVersion: v1
kind: Secret
metadata:
  annotations:
    kubernetes.io/service-account.name: my-serviceaccount
    kubernetes.io/service-account.uid: 86e98645-2e05-11e9-863a-b2d4d086dd5a
  name: my-serviceaccount-token-ntxs9
type: kubernetes.io/service-account-token
data:
  ca.crt: base64(cluster-ca-cert)
  namespace: base64(namespace)
  token: AAAA

Any attempt to regenerate the token or creating a new such secret will again make the component invalidating it.

You can opt-out of this behaviour for ServiceAccounts setting .automountServiceAccountToken=false by labeling them with token-invalidator.resources.gardener.cloud/skip=true.

In order to enable the TokenInvalidator you have to set --token-invalidator-max-concurrent-workers to a value larger than 0.

Below graphic shows an overview of the Token Invalidator for Service account secrets in the Shoot cluster. image

TokenRequestor

This controller provides the service to create and auto-renew tokens via the TokenRequest API.

It provides a functionality similar to the kubelet's Service Account Token Volume Projection. It was created to handle the special case of issuing tokens to pods that run in a different cluster than the API server they communicate with (hence, using the native token volume projection feature is not possible).

The controller differentiates between source cluster and target cluster. The source cluster hosts the gardener-resource-manager pod. Secrets in this cluster are watched and modified by the controller. The target cluster can be configured to point to another cluster. The existence of ServiceAccounts are ensured and token requests are issued against the target. When the gardener-resource-manager is deployed next to the Shoot's controlplane in the Seed the source cluster is the Seed while the target cluster points to the Shoot.

Reconciliation Loop

This controller reconciles secrets in all namespaces in the source cluster with the label: resources.gardener.cloud/purpose: token-requestor. See here for an example of the secret.

The controller ensures a ServiceAccount exists in the target cluster as specified in the annotations of the Secret in the source cluster:

serviceaccount.resources.gardener.cloud/name: <sa-name>
serviceaccount.resources.gardener.cloud/namespace: <sa-namespace>

The requested tokens will act with the privileges which are assigned to this ServiceAccount.

The controller will then request a token via the TokenRequest API and populate it into the .data.token field to the Secret in the source cluster.

Alternatively, the client can provide a raw kubeconfig (in YAML or JSON format) via the Secret's .data.kubeconfig field. The controller will then populate the requested token in the kubeconfig for the user used in the .current-context. For example, if .data.kubeconfig is

apiVersion: v1
clusters:
- cluster:
    certificate-authority-data: AAAA
    server: some-server-url
  name: shoot--foo--bar
contexts:
- context:
    cluster: shoot--foo--bar
    user: shoot--foo--bar-token
  name: shoot--foo--bar
current-context: shoot--foo--bar
kind: Config
preferences: {}
users:
- name: shoot--foo--bar-token
  user:
    token: ""

then the .users[0].user.token field of the kubeconfig will be updated accordingly.

The controller also adds an annotation to the Secret to keep track when to renew the token before it expires. By default, the tokens are issued to expire after 12 hours. The expiration time can be set with the following annotation:

serviceaccount.resources.gardener.cloud/token-expiration-duration: 6h

It automatically renews once 80% of the lifetime is reached or after 24h.

Optionally, the controller can also populate the token into a Secret in the target cluster. This can be requested by annotating the Secret in the source cluster with

token-requestor.resources.gardener.cloud/target-secret-name: "foo"
token-requestor.resources.gardener.cloud/target-secret-namespace: "bar"

Overall, the TokenRequestor controller provides credentials with limited lifetime (JWT tokens) used by Shoot control plane components running in the Seed to talk to the Shoot API Server. Please see the graphic below:

image

Webhooks

Auto-Mounting Projected ServiceAccount Tokens

When this webhook is activated then it automatically injects projected ServiceAccount token volumes into Pods and all its containers if all of the following preconditions are fulfilled:

  1. The Pod is NOT labeled with projected-token-mount.resources.gardener.cloud/skip=true.
  2. The Pod's .spec.serviceAccountName field is NOT empty and NOT set to default.
  3. The ServiceAccount specified in the Pod's .spec.serviceAccountName sets .automountServiceAccountToken=false.
  4. The Pod's .spec.volumes[] DO NOT already contain a volume with a name prefixed with kube-api-access-.

The projected volume will look as follows:

spec:
  volumes:
  - name: kube-api-access-gardener
    projected:
      defaultMode: 420
      sources:
      - serviceAccountToken:
          expirationSeconds: 43200
          path: token
      - configMap:
          items:
          - key: ca.crt
            path: ca.crt
          name: kube-root-ca.crt
      - downwardAPI:
          items:
          - fieldRef:
              apiVersion: v1
              fieldPath: metadata.namespace
            path: namespace

The expirationSeconds are defaulted to 12h and can be overwritten with the --projected-token-mount-expiration-seconds flag, or with the projected-token-mount.resources.gardener.cloud/expiration-seconds annotation on a Pod resource.

The volume will be mounted into all containers specified in the Pod to the path /var/run/secrets/kubernetes.io/serviceaccount. This is the default location where client libraries expect to find the tokens and mimics the upstream ServiceAccount admission plugin, see this document for more information.

Overall, this webhook is used to inject projected service account tokens into pods running in the Shoot and the Seed cluster. Hence, it is served from the Seed GRM and each Shoot GRM. Please find an overview below for pods deployed in the Shoot cluster:

image