Skip to content

resource-manager.md

Latest commit

 

History

History
1101 lines (887 loc) · 55.1 KB

resource-manager.md

File metadata and controls

1101 lines (887 loc) · 55.1 KB

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).

Component Configuration

Similar to other Gardener components, the gardener-resource-manager uses a so-called component configuration file. It allows specifying certain central settings like log level and formatting, client connection configuration, server ports and bind addresses, etc. In addition, controllers and webhooks can be configured and sometimes even disabled.

Note that the very basic ManagedResource, secret, and health controllers cannot be disabled.

You can find an example configuration file here.

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 the following supported modes:

  • 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.

Skipping Health Check

If a resource in the ManagedResource is annotated with resources.gardener.cloud/skip-health-check=true, then the resource will be skipped during health checks by the health controller. The ManagedResource conditions will not reflect the health condition of this resource anymore. The ResourcesProgressing condition will also be set to False.

Resource Class

By default, the gardener-resource-manager controller watches for ManagedResources in all namespaces. The .sourceClientConnection.namespace field in the component configuration restricts the watch to ManagedResources in a single namespace only. Note that this setting also affects all other controllers and webhooks since it's a central configuration.

A ManagedResource has an optional .spec.class field that allows it to indicate that it belongs to a given class of resources. The .controllers.resourceClass field in the component configuration restricts 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. The default format is <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 .controllers.clusterID field in the component configuration. 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).

By default, cluster id is not used. If cluster id is specified, the format is <cluster id>:<namespace>/<objectname>.

In addition to the origin annotation, all objects managed by the resource manager get a dedicated label resources.gardener.cloud/managed-by. This label can be used to describe these objects with a selector. By default it is set to "gardener", but this can be overwritten by setting the .conrollers.managedResources.managedByLabelValue field in the component configuration.

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 a 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 the so-called checksum annotations in the pod template, which makes Kubernetes 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 also to 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/Secrets 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, ManagedResources and for each of them:
    • iterate over the .metadata.annotations and for each of them:
      • 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".

Note: Managed resource secrets are garbage collected only if the GC controller is configured to do so. This is the case when the GC controller acts on a seed cluster.

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 setting the .controllers.garbageCollector.enabled field to true in the component configuration.

TokenInvalidator Controller

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.

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 the 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 .targetClientConnection.namespace field in the component configuration. Note that this setting also affects all other controllers and webhooks since it's a central configuration.

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 both .controllers.tokenValidator.enabled=true and .webhooks.tokenValidator.enabled=true in the component configuration.

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

TokenRequestor Controller

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 this YAML file 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

ℹ️ Generally, the controller can run with multiple instances in different components. For example, gardener-resource-manager might run the TokenRequestor controller, but gardenlet might run it, too. In order to differentiate which instance of the controller is responsible for a Secret, it can be labeled with resources.gardener.cloud/class=<class>. The <class> must be configured in the respective controller, otherwise it will be responsible for all Secrets no matter whether they have the label or not.

Kubelet Server CertificateSigningRequest Approver

Gardener configures the kubelets such that they request two certificates via the CertificateSigningRequest API:

  1. client certificate for communicating with the kube-apiserver
  2. server certificate for serving its HTTPS server

For client certificates, the kubernetes.io/kube-apiserver-client-kubelet signer is used (see Certificate Signing Requests for more details). The kube-controller-manager's csrapprover controller is responsible for auto-approving such CertificateSigningRequests so that the respective certificates can be issued.

For server certificates, the kubernetes.io/kubelet-serving signer is used. Unfortunately, the kube-controller-manager is not able to auto-approve such CertificateSigningRequests (see kubernetes/kubernetes#73356 for details).

That's the motivation for having this controller as part of gardener-resource-manager. It watches CertificateSigningRequests with the kubernetes.io/kubelet-serving signer and auto-approves them when all the following conditions are met:

  • The .spec.username is prefixed with system:node:.
  • There must be at least one DNS name or IP address as part of the certificate SANs.
  • The common name in the CSR must match the .spec.username.
  • The organization in the CSR must only contain system:nodes.
  • There must be a Node object with the same name in the shoot cluster.
  • There must be exactly one Machine for the node in the seed cluster.
  • The DNS names part of the SANs must be equal to all .status.addresses[] of type Hostname in the Node.
  • The IP addresses part of the SANs must be equal to all .status.addresses[] of type InternalIP in the Node.

If any one of these requirements is violated, the CertificateSigningRequest will be denied. Otherwise, once approved, the kube-controller-manager's csrsigner controller will issue the requested certificate.

NetworkPolicy Controller

This controller reconciles Services with a non-empty .spec.podSelector. It creates two NetworkPolicys for each port in the .spec.ports[] list. For example:

apiVersion: v1
kind: Service
metadata:
  name: gardener-resource-manager
  namespace: a
spec:
  selector:
    app: gardener-resource-manager
  ports:
  - name: server
    port: 443
    protocol: TCP
    targetPort: 10250

leads to

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  annotations:
    gardener.cloud/description: Allows ingress TCP traffic to port 10250 for pods
      selected by the a/gardener-resource-manager service selector from pods running
      in namespace a labeled with map[networking.resources.gardener.cloud/to-gardener-resource-manager-tcp-10250:allowed].
  name: ingress-to-gardener-resource-manager-tcp-10250
  namespace: a
spec:
  ingress:
  - from:
    - podSelector:
        matchLabels:
          networking.resources.gardener.cloud/to-gardener-resource-manager-tcp-10250: allowed
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      app: gardener-resource-manager
  policyTypes:
  - Ingress
---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  annotations:
    gardener.cloud/description: Allows egress TCP traffic to port 10250 from pods
      running in namespace a labeled with map[networking.resources.gardener.cloud/to-gardener-resource-manager-tcp-10250:allowed]
      to pods selected by the a/gardener-resource-manager service selector.
  name: egress-to-gardener-resource-manager-tcp-10250
  namespace: a
spec:
  egress:
  - to:
    - podSelector:
        matchLabels:
          app: gardener-resource-manager
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      networking.resources.gardener.cloud/to-gardener-resource-manager-tcp-10250: allowed
  policyTypes:
  - Egress

A component that initiates the connection to gardener-resource-manager's tcp/10250 port can now be labeled with networking.resources.gardener.cloud/to-gardener-resource-manager-tcp-10250=allowed. That's all this component needs to do - it does not need to create any NetworkPolicys itself.

Cross-Namespace Communication

Apart from this "simple" case where both communicating components run in the same namespace a, there is also the cross-namespace communication case. With above example, let's say there are components running in another namespace b, and they would like to initiate the communication with gardener-resource-manager in a. To cover this scenario, the Service can be annotated with networking.resources.gardener.cloud/namespace-selectors='[{"matchLabels":{"kubernetes.io/metadata.name":"b"}}]'.

Note that you can specify multiple namespace selectors in this annotation which are OR-ed.

This will make the controller create additional NetworkPolicys as follows:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  annotations:
    gardener.cloud/description: Allows ingress TCP traffic to port 10250 for pods selected
      by the a/gardener-resource-manager service selector from pods running in namespace b
      labeled with map[networking.resources.gardener.cloud/to-a-gardener-resource-manager-tcp-10250:allowed].
  name: ingress-to-gardener-resource-manager-tcp-10250-from-b
  namespace: a
spec:
  ingress:
  - from:
    - namespaceSelector:
        matchLabels:
          kubernetes.io/metadata.name: b
      podSelector:
        matchLabels:
          networking.resources.gardener.cloud/to-a-gardener-resource-manager-tcp-10250: allowed
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      app: gardener-resource-manager
  policyTypes:
  - Ingress
---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  annotations:
    gardener.cloud/description: Allows egress TCP traffic to port 10250 from pods running in
      namespace b labeled with map[networking.resources.gardener.cloud/to-a-gardener-resource-manager-tcp-10250:allowed]
      to pods selected by the a/gardener-resource-manager service selector.
  name: egress-to-a-gardener-resource-manager-tcp-10250
  namespace: b
spec:
  egress:
  - to:
    - namespaceSelector:
        matchLabels:
          kubernetes.io/metadata.name: a
      podSelector:
        matchLabels:
          app: gardener-resource-manager
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      networking.resources.gardener.cloud/to-a-gardener-resource-manager-tcp-10250: allowed
  policyTypes:
  - Egress

The components in namespace b now need to be labeled with networking.resources.gardener.cloud/to-a-gardener-resource-manager-tcp-10250=allowed, but that's already it.

Obviously, this approach also works for namespace selectors different from kubernetes.io/metadata.name to cover scenarios where the namespace name is not known upfront or where multiple namespaces with a similar label are relevant. The controller creates two dedicated policies for each namespace matching the selectors.

Service Targets In Multiple Namespaces

Finally, let's say there is a Service called example which exists in different namespaces whose names are not static (e.g., foo-1, foo-2), and a component in namespace bar wants to initiate connections with all of them.

The example Services in these namespaces can now be annotated with networking.resources.gardener.cloud/namespace-selectors='[{"matchLabels":{"kubernetes.io/metadata.name":"bar"}}]'. As a consequence, the component in namespace bar now needs to be labeled with networking.resources.gardener.cloud/to-foo-1-example-tcp-8080=allowed, networking.resources.gardener.cloud/to-foo-2-example-tcp-8080=allowed, etc. This approach does not work in practice, however, since the namespace names are neither static nor known upfront.

To overcome this, it is possible to specify an alias for the concrete namespace in the pod label selector via the networking.resources.gardener.cloud/pod-label-selector-namespace-alias annotation.

In above case, the example Service in the foo-* namespaces could be annotated with networking.resources.gardener.cloud/pod-label-selector-namespace-alias=all-foos. This would modify the label selector in all NetworkPolicys related to cross-namespace communication, i.e. instead of networking.resources.gardener.cloud/to-foo-{1,2,...}-example-tcp-8080=allowed, networking.resources.gardener.cloud/to-all-foos-example-tcp-8080=allowed would be used. Now the component in namespace bar only needs this single label and is able to talk to all such Services in the different namespaces.

Real-world examples for this scenario are the kube-apiserver Service (which exists in all shoot namespaces), or the istio-ingressgateway Service (which exists in all istio-ingress* namespaces). In both cases, the names of the namespaces are not statically known and depend on user input.

Overwriting The Pod Selector Label

For a component which initiates the connection to many other components, it's sometimes impractical to specify all the respective labels in its pod template. For example, let's say a component foo talks to bar{0..9} on ports tcp/808{0..9}. foo would need to have the ten networking.resources.gardener.cloud/to-bar{0..9}-tcp-808{0..9}=allowed labels.

As an alternative and to simplify this, it is also possible to annotate the targeted Services with networking.resources.gardener.cloud/from-<some-alias>-allowed-ports. For our example, <some-alias> could be all-bars.

As a result, component foo just needs to have the label networking.resources.gardener.cloud/to-all-bars=allowed instead of all the other ten explicit labels.

⚠️ Note that this also requires to specify the list of allowed container ports as annotation value since the pod selector label will no longer be specific for a dedicated service/port. For our example, the Service for barX with X in {0..9} needs to be annotated with networking.resources.gardener.cloud/from-all-bars-allowed-ports=[{"port":808X,"protocol":"TCP"}] in addition.

Real-world examples for this scenario are the Prometheis in seed clusters which initiate the communication to a lot of components in order to scrape their metrics. Another example is the kube-apiserver which initiates the communication to webhook servers (potentially of extension components that are not known by Gardener itself).

Ingress From Everywhere

All above scenarios are about components initiating connections to some targets. However, some components also receive incoming traffic from sources outside the cluster. This traffic requires adequate ingress policies so that it can be allowed.

To cover this scenario, the Service can be annotated with networking.resources.gardener.cloud/from-world-to-ports=[{"port":"10250","protocol":"TCP"}]. As a result, the controller creates the following NetworkPolicy:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: ingress-to-gardener-resource-manager-from-world
  namespace: a
spec:
  ingress:
  - from:
    - namespaceSelector: {}
      podSelector: {}
    - ipBlock:
        cidr: 0.0.0.0/0
    - ipBlock:
        cidr: ::/0
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      app: gardener-resource-manager
  policyTypes:
  - Ingress

The respective pods don't need any additional labels. If the annotation's value is empty ([]) then all ports are allowed.

Services Exposed via Ingress Resources

The controller can optionally be configured to watch Ingress resources by specifying the pod and namespace selectors for the Ingress controller. If this information is provided, it automatically creates NetworkPolicy resources allowing the respective ingress/egress traffic for the backends exposed by the Ingresses. This way, neither custom NetworkPolicys nor custom labels must be provided.

The needed configuration is part of the component configuration:

controllers:
  networkPolicy:
    enabled: true
    concurrentSyncs: 5
  # namespaceSelectors:
  # - matchLabels:
  #     kubernetes.io/metadata.name: default
    ingressControllerSelector:
      namespace: default
      podSelector:
        matchLabels:
          foo: bar

As an example, let's assume that above gardener-resource-manager Service was exposed via the following Ingress resource:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: gardener-resource-manager
  namespace: a
spec:
  rules:
  - host: grm.foo.example.com
    http:
      paths:
      - backend:
          service:
            name: gardener-resource-manager
            port:
              number: 443
        path: /
        pathType: Prefix

As a result, the controller would automatically create the following NetworkPolicys:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  annotations:
    gardener.cloud/description: Allows ingress TCP traffic to port 10250 for pods
      selected by the a/gardener-resource-manager service selector from ingress controller
      pods running in the default namespace labeled with map[foo:bar].
  name: ingress-to-gardener-resource-manager-tcp-10250-from-ingress-controller
  namespace: a
spec:
  ingress:
  - from:
    - podSelector:
        matchLabels:
          foo: bar
      namespaceSelector:
        matchLabels:
          kubernetes.io/metadata.name: default
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      app: gardener-resource-manager
  policyTypes:
  - Ingress
---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  annotations:
    gardener.cloud/description: Allows egress TCP traffic to port 10250 from pods
      running in the default namespace labeled with map[foo:bar] to pods selected by
      the a/gardener-resource-manager service selector.
  name: egress-to-a-gardener-resource-manager-tcp-10250-from-ingress-controller
  namespace: default
spec:
  egress:
  - to:
    - podSelector:
        matchLabels:
          app: gardener-resource-manager
      namespaceSelector:
        matchLabels:
          kubernetes.io/metadata.name: a
    ports:
    - port: 10250
      protocol: TCP
  podSelector:
    matchLabels:
      foo: bar
  policyTypes:
  - Egress

ℹ️ Note that Ingress resources reference the service port while NetworkPolicys reference the target port/container port. The controller automatically translates this when reconciling the NetworkPolicy resources.

Node Controller

Gardenlet configures kubelet of shoot worker nodes to register the Node object with the node.gardener.cloud/critical-components-not-ready taint (effect NoSchedule). This controller watches newly created Node objects in the shoot cluster and removes the taint once all node-critical components are scheduled and ready. If the controller finds node-critical components that are not scheduled or not ready yet, it checks the Node again after the duration configured in ResourceManagerConfiguration.controllers.node.backoff Please refer to the feature documentation or proposal issue for more details.

Webhooks

Mutating Webhooks

High Availability Config

This webhook is used to conveniently apply the configuration to make components deployed to seed or shoot clusters highly available. The details and scenarios are described in High Availability Of Deployed Components.

The webhook reacts on creation/update of Deployments, StatefulSets, HorizontalPodAutoscalers and HVPAs in namespaces labeled with high-availability-config.resources.gardener.cloud/consider=true.

The webhook performs the following actions:

  1. The .spec.replicas (or spec.minReplicas respectively) field is mutated based on the high-availability-config.resources.gardener.cloud/type label of the resource and the high-availability-config.resources.gardener.cloud/failure-tolerance-type annotation of the namespace:

    Failure Tolerance Type ➡️
    /
    ⬇️ Component Type️ ️    		 unset empty non-empty
    controller 2 1 2
    server 2 2 2
    • The replica count values can be overwritten by the high-availability-config.resources.gardener.cloud/replicas annotation.
    • It does NOT mutate the replicas when:
      • the replicas are already set to 0 (hibernation case), or
      • when the resource is scaled horizontally by HorizontalPodAutoscaler or Hvpa, and the current replica count is higher than what was computed above.

  2. When the high-availability-config.resources.gardener.cloud/zones annotation is NOT empty and either the high-availability-config.resources.gardener.cloud/failure-tolerance-type annotation is set or the high-availability-config.resources.gardener.cloud/zone-pinning annotation is set to true, then it adds a node affinity to the pod template spec:

    spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: topology.kubernetes.io/zone
            operator: In
            values:
            - <zone1>
          # - ...

    This ensures that all pods are pinned to only nodes in exactly those concrete zones.

  3. Topology Spread Constraints are added to the pod template spec when the .spec.replicas are greater than 1. When the high-availability-config.resources.gardener.cloud/zones annotation ...

    • ... contains only one zone, then the following is added:

      spec:
  topologySpreadConstraints:
  - topologyKey: kubernetes.io/hostname
    maxSkew: 1
    whenUnsatisfiable: ScheduleAnyway
    labelSelector: ...

      This ensures that the (multiple) pods are scheduled across nodes on best-effort basis.

    • ... contains at least two zones, then the following is added:

      spec:
  topologySpreadConstraints:
  - topologyKey: kubernetes.io/hostname
    maxSkew: 1
    whenUnsatisfiable: ScheduleAnyway
    labelSelector: ...
  - topologyKey: topology.kubernetes.io/zone
    minDomains: 2 # lower value of max replicas or number of zones
    maxSkew: 1
    whenUnsatisfiable: DoNotSchedule
    labelSelector: ...

      This enforces that the (multiple) pods are scheduled across zones. It circumvents a known limitation in Kubernetes for clusters < 1.26 (ref kubernetes/kubernetes#109364. In case the number of replicas is larger than twice the number of zones, then the maxSkew=2 for the second spread constraints. The minDomains calculation is based on whatever value is lower - (maximum) replicas or number of zones. This is the number of minimum domains required to schedule pods in a highly available manner.

    Independent on the number of zones, when the high-availability-config.resources.gardener.cloud/failure-tolerance-type annotation is set and NOT empty, then the whenUnsatisfiable is set to DoNotSchedule for the constraint with topologyKey=kubernetes.io/hostname (which enforces the node-spread).

  4. Adds default tolerations for taint-based evictions:

    Tolerations for taints node.kubernetes.io/not-ready and node.kubernetes.io/unreachable are added to the handled Deployment and StatefulSet if their podTemplates do not already specify them. The TolerationSeconds are taken from the respective configuration section of the webhook's configuration (see example)).

    We consider fine-tuned values for those tolerations a matter of high-availability because they often help to reduce recovery times in case of node or zone outages, also see High-Availability Best Practices. In addition, this webhook handling helps to set defaults for many but not all workload components in a cluster. For instance, Gardener can use this webhook to set defaults for nearly every component in seed clusters but only for the system components in shoot clusters. Any customer workload remains unchanged.

Kubernetes Service Host Injection

By default, when Pods are created, Kubernetes implicitly injects the KUBERNETES_SERVICE_HOST environment variable into all containers. The value of this variable points it to the default Kubernetes service (i.e., kubernetes.default.svc.cluster.local). This allows pods to conveniently talk to the API server of their cluster.

In shoot clusters, this network path involves the apiserver-proxy DaemonSet which eventually forwards the traffic to the API server. Hence, it results in additional network hop.

The purpose of this webhook is to explicitly inject the KUBERNETES_SERVICE_HOST environment variable into all containers and setting its value to the FQDN of the API server. This way, the additional network hop is avoided.

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 .webhooks.projectedTokenMount.expirationSeconds field in the component configuration, 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 Managing Service Accounts 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

Pod Topology Spread Constraints

When this webhook is enabled, then it mimics the topologyKey feature for Topology Spread Constraints (TSC) on the label pod-template-hash. Concretely, when a pod is labelled with pod-template-hash, the handler of this webhook extends any topology spread constraint in the pod:

metadata:
  labels:
    pod-template-hash: 123abc
spec:
  topologySpreadConstraints:
  - maxSkew: 1
    topologyKey: topology.kubernetes.io/zone
    whenUnsatisfiable: DoNotSchedule
    labelSelector:
      matchLabels:
        pod-template-hash: 123abc # added by webhook

The procedure circumvents a known limitation with TSCs which leads to imbalanced deployments after rolling updates. Gardener enables this webhook to schedule pods of deployments across nodes and zones.

Please note that the gardener-resource-manager itself as well as pods labelled with topology-spread-constraints.resources.gardener.cloud/skip are excluded from any mutations.

System Components Webhook

If enabled, this webhook handles scheduling concerns for system components Pods (except those managed by DaemonSets). The following tasks are performed by this webhook:

  1. Add pod.spec.nodeSelector as given in the webhook configuration.
  2. Add pod.spec.tolerations as given in the webhook configuration.
  3. Add pod.spec.tolerations for any existing nodes matching the node selector given in the webhook configuration. Known taints and tolerations used for taint based evictions are disregarded.

Gardener enables this webhook for kube-system and kubernetes-dashboard namespaces in shoot clusters, selecting Pods being labelled with resources.gardener.cloud/managed-by: gardener. It adds a configuration, so that Pods will get the worker.gardener.cloud/system-components: true node selector (step 1) as well as tolerate any custom taint (step 2) that is added to system component worker nodes (shoot.spec.provider.workers[].systemComponents.allow: true). In addition, the webhook merges these tolerations with the ones required for at that time available system component Nodes in the cluster (step 3). Both is required to ensure system component Pods can be scheduled or executed during an active shoot reconciliation that is happening due to any modifications to shoot.spec.provider.workers[].taints, e.g. Pods must be scheduled while there are still Nodes not having the updated taint configuration.

You can opt-out of this behaviour for Pods by labeling them with system-components-config.resources.gardener.cloud/skip=true.

EndpointSlice Hints

This webhook mutates EndpointSlices. For each endpoint in the EndpointSlice, it sets the endpoint's hints to the endpoint's zone.

apiVersion: discovery.k8s.io/v1
kind: EndpointSlice
metadata:
  name: example-hints
endpoints:
- addresses:
  - "10.1.2.3"
  conditions:
    ready: true
  hostname: pod-1
  zone: zone-a
  hints:
    forZones:
    - name: "zone-a" # added by webhook
- addresses:
  - "10.1.2.4"
  conditions:
    ready: true
  hostname: pod-2
  zone: zone-b
  hints:
    forZones:
    - name: "zone-b" # added by webhook

The webhook aims to circumvent issues with the Kubernetes TopologyAwareHints feature that currently does not allow to achieve a deterministic topology-aware traffic routing. For more details, see the following issue kubernetes/kubernetes#113731 that describes drawbacks of the TopologyAwareHints feature for our use case. If the above-mentioned issue gets resolved and there is a native support for deterministic topology-aware traffic routing in Kubernetes, then this webhook can be dropped in favor of the native Kubernetes feature.

Validating Webhooks

Unconfirmed Deletion Prevention For Custom Resources And Definitions

As part of Gardener's extensibility concepts, a lot of CustomResourceDefinitions are deployed to the seed clusters that serve as extension points for provider-specific controllers. For example, the Infrastructure CRD triggers the provider extension to prepare the IaaS infrastructure of the underlying cloud provider for a to-be-created shoot cluster. Consequently, these extension CRDs have a lot of power and control large portions of the end-user's shoot cluster. Accidental or undesired deletions of those resource can cause tremendous and hard-to-recover-from outages and should be prevented.

When this webhook is activated, it reacts for CustomResourceDefinitions and most of the custom resources in the extensions.gardener.cloud/v1alpha1 API group. It also reacts for the druid.gardener.cloud/v1alpha1.Etcd resources.

The webhook prevents DELETE requests for those CustomResourceDefinitions labeled with gardener.cloud/deletion-protected=true, and for all mentioned custom resources if they were not previously annotated with the confirmation.gardener.cloud/deletion=true. This prevents that undesired kubectl delete <...> requests are accepted.

Extension Resource Validation

When this webhook is activated, it reacts for most of the custom resources in the extensions.gardener.cloud/v1alpha1 API group. It also reacts for the druid.gardener.cloud/v1alpha1.Etcd resources.

The webhook validates the resources specifications for CREATE and UPDATE requests.