Cluster Architecture

• 1: Nodes
• 2: Communication between Nodes and the Control Plane
• 3: Controllers
• 4: Leases
• 5: Cloud Controller Manager
• 6: About cgroup v2
• 7: Container Runtime Interface (CRI)
• 8: Garbage Collection
• 9: Mixed Version Proxy

1 - Nodes

Kubernetes runs your workload by placing containers into Pods to run on Nodes. A node may be a virtual or physical machine, depending on the cluster. Each node is managed by the control plane and contains the services necessary to run Pods.

Typically you have several nodes in a cluster; in a learning or resource-limited environment, you might have only one node.

The components on a node include the kubelet, a container runtime, and the kube-proxy.

Management

There are two main ways to have Nodes added to the API server:

1. The kubelet on a node self-registers to the control plane
2. You (or another human user) manually add a Node object

After you create a Node object, or the kubelet on a node self-registers, the control plane checks whether the new Node object is valid. For example, if you try to create a Node from the following JSON manifest:

{
  "kind": "Node",
  "apiVersion": "v1",
  "metadata": {
    "name": "10.240.79.157",
    "labels": {
      "name": "my-first-k8s-node"
    }
  }
}

Kubernetes creates a Node object internally (the representation). Kubernetes checks that a kubelet has registered to the API server that matches the metadata.name field of the Node. If the node is healthy (i.e. all necessary services are running), then it is eligible to run a Pod. Otherwise, that node is ignored for any cluster activity until it becomes healthy.

The name of a Node object must be a valid DNS subdomain name.

Node name uniqueness

The name identifies a Node. Two Nodes cannot have the same name at the same time. Kubernetes also assumes that a resource with the same name is the same object. In case of a Node, it is implicitly assumed that an instance using the same name will have the same state (e.g. network settings, root disk contents) and attributes like node labels. This may lead to inconsistencies if an instance was modified without changing its name. If the Node needs to be replaced or updated significantly, the existing Node object needs to be removed from API server first and re-added after the update.

Self-registration of Nodes

When the kubelet flag --register-node is true (the default), the kubelet will attempt to register itself with the API server. This is the preferred pattern, used by most distros.

For self-registration, the kubelet is started with the following options:

• --kubeconfig - Path to credentials to authenticate itself to the API server.

• --cloud-provider - How to talk to a cloud provider to read metadata about itself.

• --register-node - Automatically register with the API server.

• --register-with-taints - Register the node with the given list of taints (comma separated <key>=<value>:<effect>).

  No-op if register-node is false.

• --node-ip - Optional comma-separated list of the IP addresses for the node. You can only specify a single address for each address family. For example, in a single-stack IPv4 cluster, you set this value to be the IPv4 address that the kubelet should use for the node. See configure IPv4/IPv6 dual stack for details of running a dual-stack cluster.

  If you don't provide this argument, the kubelet uses the node's default IPv4 address, if any; if the node has no IPv4 addresses then the kubelet uses the node's default IPv6 address.

• --node-labels - Labels to add when registering the node in the cluster (see label restrictions enforced by the NodeRestriction admission plugin).

• --node-status-update-frequency - Specifies how often kubelet posts its node status to the API server.

When the Node authorization mode and NodeRestriction admission plugin are enabled, kubelets are only authorized to create/modify their own Node resource.

Manual Node administration

You can create and modify Node objects using kubectl.

When you want to create Node objects manually, set the kubelet flag --register-node=false.

You can modify Node objects regardless of the setting of --register-node. For example, you can set labels on an existing Node or mark it unschedulable.

You can use labels on Nodes in conjunction with node selectors on Pods to control scheduling. For example, you can constrain a Pod to only be eligible to run on a subset of the available nodes.

Marking a node as unschedulable prevents the scheduler from placing new pods onto that Node but does not affect existing Pods on the Node. This is useful as a preparatory step before a node reboot or other maintenance.

To mark a Node unschedulable, run:

kubectl cordon $NODENAME

See Safely Drain a Node for more details.

Node status

A Node's status contains the following information:

• Addresses
• Conditions
• Capacity and Allocatable
• Info

You can use kubectl to view a Node's status and other details:

kubectl describe node <insert-node-name-here>

See Node Status for more details.

Node heartbeats

Heartbeats, sent by Kubernetes nodes, help your cluster determine the availability of each node, and to take action when failures are detected.

For nodes there are two forms of heartbeats:

• Updates to the .status of a Node.
• Lease objects within the kube-node-lease namespace. Each Node has an associated Lease object.

Node controller

The node controller is a Kubernetes control plane component that manages various aspects of nodes.

The node controller has multiple roles in a node's life. The first is assigning a CIDR block to the node when it is registered (if CIDR assignment is turned on).

The second is keeping the node controller's internal list of nodes up to date with the cloud provider's list of available machines. When running in a cloud environment and whenever a node is unhealthy, the node controller asks the cloud provider if the VM for that node is still available. If not, the node controller deletes the node from its list of nodes.

The third is monitoring the nodes' health. The node controller is responsible for:

• In the case that a node becomes unreachable, updating the Ready condition in the Node's .status field. In this case the node controller sets the Ready condition to Unknown.
• If a node remains unreachable: triggering API-initiated eviction for all of the Pods on the unreachable node. By default, the node controller waits 5 minutes between marking the node as Unknown and submitting the first eviction request.

By default, the node controller checks the state of each node every 5 seconds. This period can be configured using the --node-monitor-period flag on the kube-controller-manager component.

Rate limits on eviction

In most cases, the node controller limits the eviction rate to --node-eviction-rate (default 0.1) per second, meaning it won't evict pods from more than 1 node per 10 seconds.

The node eviction behavior changes when a node in a given availability zone becomes unhealthy. The node controller checks what percentage of nodes in the zone are unhealthy (the Ready condition is Unknown or False) at the same time:

• If the fraction of unhealthy nodes is at least --unhealthy-zone-threshold (default 0.55), then the eviction rate is reduced.
• If the cluster is small (i.e. has less than or equal to --large-cluster-size-threshold nodes - default 50), then evictions are stopped.
• Otherwise, the eviction rate is reduced to --secondary-node-eviction-rate (default 0.01) per second.

The reason these policies are implemented per availability zone is because one availability zone might become partitioned from the control plane while the others remain connected. If your cluster does not span multiple cloud provider availability zones, then the eviction mechanism does not take per-zone unavailability into account.

A key reason for spreading your nodes across availability zones is so that the workload can be shifted to healthy zones when one entire zone goes down. Therefore, if all nodes in a zone are unhealthy, then the node controller evicts at the normal rate of --node-eviction-rate. The corner case is when all zones are completely unhealthy (none of the nodes in the cluster are healthy). In such a case, the node controller assumes that there is some problem with connectivity between the control plane and the nodes, and doesn't perform any evictions. (If there has been an outage and some nodes reappear, the node controller does evict pods from the remaining nodes that are unhealthy or unreachable).

The node controller is also responsible for evicting pods running on nodes with NoExecute taints, unless those pods tolerate that taint. The node controller also adds taints corresponding to node problems like node unreachable or not ready. This means that the scheduler won't place Pods onto unhealthy nodes.

Resource capacity tracking

Node objects track information about the Node's resource capacity: for example, the amount of memory available and the number of CPUs. Nodes that self register report their capacity during registration. If you manually add a Node, then you need to set the node's capacity information when you add it.

The Kubernetes scheduler ensures that there are enough resources for all the Pods on a Node. The scheduler checks that the sum of the requests of containers on the node is no greater than the node's capacity. That sum of requests includes all containers managed by the kubelet, but excludes any containers started directly by the container runtime, and also excludes any processes running outside of the kubelet's control.

Node topology

If you have enabled the TopologyManager feature gate, then the kubelet can use topology hints when making resource assignment decisions. See Control Topology Management Policies on a Node for more information.

Swap memory management

To enable swap on a node, the NodeSwap feature gate must be enabled on the kubelet (default is true), and the --fail-swap-on command line flag or failSwapOn configuration setting must be set to false. To allow Pods to utilize swap, swapBehavior should not be set to NoSwap (which is the default behavior) in the kubelet config.

A user can also optionally configure memorySwap.swapBehavior in order to specify how a node will use swap memory. For example,

memorySwap:
  swapBehavior: LimitedSwap

• NoSwap (default): Kubernetes workloads will not use swap.
• LimitedSwap: The utilization of swap memory by Kubernetes workloads is subject to limitations. Only Pods of Burstable QoS are permitted to employ swap.

If configuration for memorySwap is not specified and the feature gate is enabled, by default the kubelet will apply the same behaviour as the NoSwap setting.

With LimitedSwap, Pods that do not fall under the Burstable QoS classification (i.e. BestEffort/Guaranteed Qos Pods) are prohibited from utilizing swap memory. To maintain the aforementioned security and node health guarantees, these Pods are not permitted to use swap memory when LimitedSwap is in effect.

Prior to detailing the calculation of the swap limit, it is necessary to define the following terms:

• nodeTotalMemory: The total amount of physical memory available on the node.
• totalPodsSwapAvailable: The total amount of swap memory on the node that is available for use by Pods (some swap memory may be reserved for system use).
• containerMemoryRequest: The container's memory request.

Swap limitation is configured as: (containerMemoryRequest / nodeTotalMemory) * totalPodsSwapAvailable.

It is important to note that, for containers within Burstable QoS Pods, it is possible to opt-out of swap usage by specifying memory requests that are equal to memory limits. Containers configured in this manner will not have access to swap memory.

Swap is supported only with cgroup v2, cgroup v1 is not supported.

For more information, and to assist with testing and provide feedback, please see the blog-post about Kubernetes 1.28: NodeSwap graduates to Beta1, KEP-2400 and its design proposal.

What's next

Learn more about the following:

• Components that make up a node.
• API definition for Node.
• Node section of the architecture design document.
• Graceful/non-graceful node shutdown.
• Cluster autoscaling to manage the number and size of nodes in your cluster.
• Taints and Tolerations.
• Node Resource Managers.
• Resource Management for Windows nodes.

2 - Communication between Nodes and the Control Plane

This document catalogs the communication paths between the API server and the Kubernetes cluster. The intent is to allow users to customize their installation to harden the network configuration such that the cluster can be run on an untrusted network (or on fully public IPs on a cloud provider).

Node to Control Plane

Kubernetes has a "hub-and-spoke" API pattern. All API usage from nodes (or the pods they run) terminates at the API server. None of the other control plane components are designed to expose remote services. The API server is configured to listen for remote connections on a secure HTTPS port (typically 443) with one or more forms of client authentication enabled. One or more forms of authorization should be enabled, especially if anonymous requests or service account tokens are allowed.

Nodes should be provisioned with the public root certificate for the cluster such that they can connect securely to the API server along with valid client credentials. A good approach is that the client credentials provided to the kubelet are in the form of a client certificate. See kubelet TLS bootstrapping for automated provisioning of kubelet client certificates.

Pods that wish to connect to the API server can do so securely by leveraging a service account so that Kubernetes will automatically inject the public root certificate and a valid bearer token into the pod when it is instantiated. The kubernetes service (in default namespace) is configured with a virtual IP address that is redirected (via kube-proxy) to the HTTPS endpoint on the API server.

The control plane components also communicate with the API server over the secure port.

As a result, the default operating mode for connections from the nodes and pod running on the nodes to the control plane is secured by default and can run over untrusted and/or public networks.

Control plane to node

There are two primary communication paths from the control plane (the API server) to the nodes. The first is from the API server to the kubelet process which runs on each node in the cluster. The second is from the API server to any node, pod, or service through the API server's proxy functionality.

API server to kubelet

The connections from the API server to the kubelet are used for:

• Fetching logs for pods.
• Attaching (usually through kubectl) to running pods.
• Providing the kubelet's port-forwarding functionality.

These connections terminate at the kubelet's HTTPS endpoint. By default, the API server does not verify the kubelet's serving certificate, which makes the connection subject to man-in-the-middle attacks and unsafe to run over untrusted and/or public networks.

To verify this connection, use the --kubelet-certificate-authority flag to provide the API server with a root certificate bundle to use to verify the kubelet's serving certificate.

If that is not possible, use SSH tunneling between the API server and kubelet if required to avoid connecting over an untrusted or public network.

Finally, Kubelet authentication and/or authorization should be enabled to secure the kubelet API.

API server to nodes, pods, and services

The connections from the API server to a node, pod, or service default to plain HTTP connections and are therefore neither authenticated nor encrypted. They can be run over a secure HTTPS connection by prefixing https: to the node, pod, or service name in the API URL, but they will not validate the certificate provided by the HTTPS endpoint nor provide client credentials. So while the connection will be encrypted, it will not provide any guarantees of integrity. These connections are not currently safe to run over untrusted or public networks.

SSH tunnels

Kubernetes supports SSH tunnels to protect the control plane to nodes communication paths. In this configuration, the API server initiates an SSH tunnel to each node in the cluster (connecting to the SSH server listening on port 22) and passes all traffic destined for a kubelet, node, pod, or service through the tunnel. This tunnel ensures that the traffic is not exposed outside of the network in which the nodes are running.

Konnectivity service

As a replacement to the SSH tunnels, the Konnectivity service provides TCP level proxy for the control plane to cluster communication. The Konnectivity service consists of two parts: the Konnectivity server in the control plane network and the Konnectivity agents in the nodes network. The Konnectivity agents initiate connections to the Konnectivity server and maintain the network connections. After enabling the Konnectivity service, all control plane to nodes traffic goes through these connections.

Follow the Konnectivity service task to set up the Konnectivity service in your cluster.

What's next

• Read about the Kubernetes control plane components
• Learn more about Hubs and Spoke model
• Learn how to Secure a Cluster
• Learn more about the Kubernetes API
• Set up Konnectivity service
• Use Port Forwarding to Access Applications in a Cluster
• Learn how to Fetch logs for Pods, use kubectl port-forward

3 - Controllers

In robotics and automation, a control loop is a non-terminating loop that regulates the state of a system.

Here is one example of a control loop: a thermostat in a room.

When you set the temperature, that's telling the thermostat about your desired state. The actual room temperature is the current state. The thermostat acts to bring the current state closer to the desired state, by turning equipment on or off.

Controller pattern

A controller tracks at least one Kubernetes resource type. These objects have a spec field that represents the desired state. The controller(s) for that resource are responsible for making the current state come closer to that desired state.

The controller might carry the action out itself; more commonly, in Kubernetes, a controller will send messages to the API server that have useful side effects. You'll see examples of this below.

Control via API server

The Job controller is an example of a Kubernetes built-in controller. Built-in controllers manage state by interacting with the cluster API server.

Job is a Kubernetes resource that runs a Pod, or perhaps several Pods, to carry out a task and then stop.

(Once scheduled, Pod objects become part of the desired state for a kubelet).

When the Job controller sees a new task it makes sure that, somewhere in your cluster, the kubelets on a set of Nodes are running the right number of Pods to get the work done. The Job controller does not run any Pods or containers itself. Instead, the Job controller tells the API server to create or remove Pods. Other components in the control plane act on the new information (there are new Pods to schedule and run), and eventually the work is done.

After you create a new Job, the desired state is for that Job to be completed. The Job controller makes the current state for that Job be nearer to your desired state: creating Pods that do the work you wanted for that Job, so that the Job is closer to completion.

Controllers also update the objects that configure them. For example: once the work is done for a Job, the Job controller updates that Job object to mark it Finished.

(This is a bit like how some thermostats turn a light off to indicate that your room is now at the temperature you set).

Direct control

In contrast with Job, some controllers need to make changes to things outside of your cluster.

For example, if you use a control loop to make sure there are enough Nodes in your cluster, then that controller needs something outside the current cluster to set up new Nodes when needed.

Controllers that interact with external state find their desired state from the API server, then communicate directly with an external system to bring the current state closer in line.

(There actually is a controller that horizontally scales the nodes in your cluster.)

The important point here is that the controller makes some changes to bring about your desired state, and then reports the current state back to your cluster's API server. Other control loops can observe that reported data and take their own actions.

In the thermostat example, if the room is very cold then a different controller might also turn on a frost protection heater. With Kubernetes clusters, the control plane indirectly works with IP address management tools, storage services, cloud provider APIs, and other services by extending Kubernetes to implement that.

Desired versus current state

Kubernetes takes a cloud-native view of systems, and is able to handle constant change.

Your cluster could be changing at any point as work happens and control loops automatically fix failures. This means that, potentially, your cluster never reaches a stable state.

As long as the controllers for your cluster are running and able to make useful changes, it doesn't matter if the overall state is stable or not.

Design

As a tenet of its design, Kubernetes uses lots of controllers that each manage a particular aspect of cluster state. Most commonly, a particular control loop (controller) uses one kind of resource as its desired state, and has a different kind of resource that it manages to make that desired state happen. For example, a controller for Jobs tracks Job objects (to discover new work) and Pod objects (to run the Jobs, and then to see when the work is finished). In this case something else creates the Jobs, whereas the Job controller creates Pods.

It's useful to have simple controllers rather than one, monolithic set of control loops that are interlinked. Controllers can fail, so Kubernetes is designed to allow for that.

Ways of running controllers

Kubernetes comes with a set of built-in controllers that run inside the kube-controller-manager. These built-in controllers provide important core behaviors.

The Deployment controller and Job controller are examples of controllers that come as part of Kubernetes itself ("built-in" controllers). Kubernetes lets you run a resilient control plane, so that if any of the built-in controllers were to fail, another part of the control plane will take over the work.

You can find controllers that run outside the control plane, to extend Kubernetes. Or, if you want, you can write a new controller yourself. You can run your own controller as a set of Pods, or externally to Kubernetes. What fits best will depend on what that particular controller does.

What's next

• Read about the Kubernetes control plane
• Discover some of the basic Kubernetes objects
• Learn more about the Kubernetes API
• If you want to write your own controller, see Kubernetes extension patterns and the sample-controller repository.

4 - Leases

Distributed systems often have a need for leases, which provide a mechanism to lock shared resources and coordinate activity between members of a set. In Kubernetes, the lease concept is represented by Lease objects in the coordination.k8s.io API Group, which are used for system-critical capabilities such as node heartbeats and component-level leader election.

Node heartbeats

Kubernetes uses the Lease API to communicate kubelet node heartbeats to the Kubernetes API server. For every Node , there is a Lease object with a matching name in the kube-node-lease namespace. Under the hood, every kubelet heartbeat is an update request to this Lease object, updating the spec.renewTime field for the Lease. The Kubernetes control plane uses the time stamp of this field to determine the availability of this Node.

See Node Lease objects for more details.

Leader election

Kubernetes also uses Leases to ensure only one instance of a component is running at any given time. This is used by control plane components like kube-controller-manager and kube-scheduler in HA configurations, where only one instance of the component should be actively running while the other instances are on stand-by.

API server identity

Starting in Kubernetes v1.26, each kube-apiserver uses the Lease API to publish its identity to the rest of the system. While not particularly useful on its own, this provides a mechanism for clients to discover how many instances of kube-apiserver are operating the Kubernetes control plane. Existence of kube-apiserver leases enables future capabilities that may require coordination between each kube-apiserver.

You can inspect Leases owned by each kube-apiserver by checking for lease objects in the kube-system namespace with the name kube-apiserver-<sha256-hash>. Alternatively you can use the label selector apiserver.kubernetes.io/identity=kube-apiserver:

kubectl -n kube-system get lease -l apiserver.kubernetes.io/identity=kube-apiserver

NAME                                        HOLDER                                                                           AGE
apiserver-07a5ea9b9b072c4a5f3d1c3702        apiserver-07a5ea9b9b072c4a5f3d1c3702_0c8914f7-0f35-440e-8676-7844977d3a05        5m33s
apiserver-7be9e061c59d368b3ddaf1376e        apiserver-7be9e061c59d368b3ddaf1376e_84f2a85d-37c1-4b14-b6b9-603e62e4896f        4m23s
apiserver-1dfef752bcb36637d2763d1868        apiserver-1dfef752bcb36637d2763d1868_c5ffa286-8a9a-45d4-91e7-61118ed58d2e        4m43s

The SHA256 hash used in the lease name is based on the OS hostname as seen by that API server. Each kube-apiserver should be configured to use a hostname that is unique within the cluster. New instances of kube-apiserver that use the same hostname will take over existing Leases using a new holder identity, as opposed to instantiating new Lease objects. You can check the hostname used by kube-apisever by checking the value of the kubernetes.io/hostname label:

kubectl -n kube-system get lease apiserver-07a5ea9b9b072c4a5f3d1c3702 -o yaml

apiVersion: coordination.k8s.io/v1
kind: Lease
metadata:
  creationTimestamp: "2023-07-02T13:16:48Z"
  labels:
    apiserver.kubernetes.io/identity: kube-apiserver
    kubernetes.io/hostname: master-1
  name: apiserver-07a5ea9b9b072c4a5f3d1c3702
  namespace: kube-system
  resourceVersion: "334899"
  uid: 90870ab5-1ba9-4523-b215-e4d4e662acb1
spec:
  holderIdentity: apiserver-07a5ea9b9b072c4a5f3d1c3702_0c8914f7-0f35-440e-8676-7844977d3a05
  leaseDurationSeconds: 3600
  renewTime: "2023-07-04T21:58:48.065888Z"

Expired leases from kube-apiservers that no longer exist are garbage collected by new kube-apiservers after 1 hour.

You can disable API server identity leases by disabling the APIServerIdentity feature gate.

Workloads

Your own workload can define its own use of Leases. For example, you might run a custom controller where a primary or leader member performs operations that its peers do not. You define a Lease so that the controller replicas can select or elect a leader, using the Kubernetes API for coordination. If you do use a Lease, it's a good practice to define a name for the Lease that is obviously linked to the product or component. For example, if you have a component named Example Foo, use a Lease named example-foo.

If a cluster operator or another end user could deploy multiple instances of a component, select a name prefix and pick a mechanism (such as hash of the name of the Deployment) to avoid name collisions for the Leases.

You can use another approach so long as it achieves the same outcome: different software products do not conflict with one another.

5 - Cloud Controller Manager

Cloud infrastructure technologies let you run Kubernetes on public, private, and hybrid clouds. Kubernetes believes in automated, API-driven infrastructure without tight coupling between components.

The cloud-controller-manager is a Kubernetes control plane component that embeds cloud-specific control logic. The cloud controller manager lets you link your cluster into your cloud provider's API, and separates out the components that interact with that cloud platform from components that only interact with your cluster.

By decoupling the interoperability logic between Kubernetes and the underlying cloud infrastructure, the cloud-controller-manager component enables cloud providers to release features at a different pace compared to the main Kubernetes project.

The cloud-controller-manager is structured using a plugin mechanism that allows different cloud providers to integrate their platforms with Kubernetes.

Design

The cloud controller manager runs in the control plane as a replicated set of processes (usually, these are containers in Pods). Each cloud-controller-manager implements multiple controllers in a single process.

Cloud controller manager functions

The controllers inside the cloud controller manager include:

Node controller

The node controller is responsible for updating Node objects when new servers are created in your cloud infrastructure. The node controller obtains information about the hosts running inside your tenancy with the cloud provider. The node controller performs the following functions:

1. Update a Node object with the corresponding server's unique identifier obtained from the cloud provider API.
2. Annotating and labelling the Node object with cloud-specific information, such as the region the node is deployed into and the resources (CPU, memory, etc) that it has available.
3. Obtain the node's hostname and network addresses.
4. Verifying the node's health. In case a node becomes unresponsive, this controller checks with your cloud provider's API to see if the server has been deactivated / deleted / terminated. If the node has been deleted from the cloud, the controller deletes the Node object from your Kubernetes cluster.

Some cloud provider implementations split this into a node controller and a separate node lifecycle controller.

Route controller

The route controller is responsible for configuring routes in the cloud appropriately so that containers on different nodes in your Kubernetes cluster can communicate with each other.

Depending on the cloud provider, the route controller might also allocate blocks of IP addresses for the Pod network.

Service controller

Services integrate with cloud infrastructure components such as managed load balancers, IP addresses, network packet filtering, and target health checking. The service controller interacts with your cloud provider's APIs to set up load balancers and other infrastructure components when you declare a Service resource that requires them.

Authorization

This section breaks down the access that the cloud controller manager requires on various API objects, in order to perform its operations.

Node controller

The Node controller only works with Node objects. It requires full access to read and modify Node objects.

v1/Node:

• get
• list
• create
• update
• patch
• watch
• delete

Route controller

The route controller listens to Node object creation and configures routes appropriately. It requires Get access to Node objects.

v1/Node:

• get

Service controller

The service controller watches for Service object create, update and delete events and then configures Endpoints for those Services appropriately (for EndpointSlices, the kube-controller-manager manages these on demand).

To access Services, it requires list, and watch access. To update Services, it requires patch and update access.

To set up Endpoints resources for the Services, it requires access to create, list, get, watch, and update.

v1/Service:

• list
• get
• watch
• patch
• update

Others

The implementation of the core of the cloud controller manager requires access to create Event objects, and to ensure secure operation, it requires access to create ServiceAccounts.

v1/Event:

• create
• patch
• update

v1/ServiceAccount:

• create

The RBAC ClusterRole for the cloud controller manager looks like:

apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRole
metadata:
  name: cloud-controller-manager
rules:
- apiGroups:
  - ""
  resources:
  - events
  verbs:
  - create
  - patch
  - update
- apiGroups:
  - ""
  resources:
  - nodes
  verbs:
  - '*'
- apiGroups:
  - ""
  resources:
  - nodes/status
  verbs:
  - patch
- apiGroups:
  - ""
  resources:
  - services
  verbs:
  - list
  - patch
  - update
  - watch
- apiGroups:
  - ""
  resources:
  - serviceaccounts
  verbs:
  - create
- apiGroups:
  - ""
  resources:
  - persistentvolumes
  verbs:
  - get
  - list
  - update
  - watch
- apiGroups:
  - ""
  resources:
  - endpoints
  verbs:
  - create
  - get
  - list
  - watch
  - update

What's next

• Cloud Controller Manager Administration has instructions on running and managing the cloud controller manager.

• To upgrade a HA control plane to use the cloud controller manager, see Migrate Replicated Control Plane To Use Cloud Controller Manager.

• Want to know how to implement your own cloud controller manager, or extend an existing project?

  • The cloud controller manager uses Go interfaces, specifically, CloudProvider interface defined in cloud.go from kubernetes/cloud-provider to allow implementations from any cloud to be plugged in.
  • The implementation of the shared controllers highlighted in this document (Node, Route, and Service), and some scaffolding along with the shared cloudprovider interface, is part of the Kubernetes core. Implementations specific to cloud providers are outside the core of Kubernetes and implement the CloudProvider interface.
  • For more information about developing plugins, see Developing Cloud Controller Manager.

6 - About cgroup v2

On Linux, control groups constrain resources that are allocated to processes.

The kubelet and the underlying container runtime need to interface with cgroups to enforce resource management for pods and containers which includes cpu/memory requests and limits for containerized workloads.

There are two versions of cgroups in Linux: cgroup v1 and cgroup v2. cgroup v2 is the new generation of the cgroup API.

What is cgroup v2?

cgroup v2 is the next version of the Linux cgroup API. cgroup v2 provides a unified control system with enhanced resource management capabilities.

cgroup v2 offers several improvements over cgroup v1, such as the following:

• Single unified hierarchy design in API
• Safer sub-tree delegation to containers
• Newer features like Pressure Stall Information
• Enhanced resource allocation management and isolation across multiple resources
  • Unified accounting for different types of memory allocations (network memory, kernel memory, etc)
  • Accounting for non-immediate resource changes such as page cache write backs

Some Kubernetes features exclusively use cgroup v2 for enhanced resource management and isolation. For example, the MemoryQoS feature improves memory QoS and relies on cgroup v2 primitives.

Using cgroup v2

The recommended way to use cgroup v2 is to use a Linux distribution that enables and uses cgroup v2 by default.

To check if your distribution uses cgroup v2, refer to Identify cgroup version on Linux nodes.

Requirements

cgroup v2 has the following requirements:

• OS distribution enables cgroup v2
• Linux Kernel version is 5.8 or later
• Container runtime supports cgroup v2. For example:
  • containerd v1.4 and later
  • cri-o v1.20 and later
• The kubelet and the container runtime are configured to use the systemd cgroup driver

Linux Distribution cgroup v2 support

For a list of Linux distributions that use cgroup v2, refer to the cgroup v2 documentation

• Container Optimized OS (since M97)
• Ubuntu (since 21.10, 22.04+ recommended)
• Debian GNU/Linux (since Debian 11 bullseye)
• Fedora (since 31)
• Arch Linux (since April 2021)
• RHEL and RHEL-like distributions (since 9)

To check if your distribution is using cgroup v2, refer to your distribution's documentation or follow the instructions in Identify the cgroup version on Linux nodes.

You can also enable cgroup v2 manually on your Linux distribution by modifying the kernel cmdline boot arguments. If your distribution uses GRUB, systemd.unified_cgroup_hierarchy=1 should be added in GRUB_CMDLINE_LINUX under /etc/default/grub, followed by sudo update-grub. However, the recommended approach is to use a distribution that already enables cgroup v2 by default.

Migrating to cgroup v2

To migrate to cgroup v2, ensure that you meet the requirements, then upgrade to a kernel version that enables cgroup v2 by default.

The kubelet automatically detects that the OS is running on cgroup v2 and performs accordingly with no additional configuration required.

There should not be any noticeable difference in the user experience when switching to cgroup v2, unless users are accessing the cgroup file system directly, either on the node or from within the containers.

cgroup v2 uses a different API than cgroup v1, so if there are any applications that directly access the cgroup file system, they need to be updated to newer versions that support cgroup v2. For example:

• Some third-party monitoring and security agents may depend on the cgroup filesystem. Update these agents to versions that support cgroup v2.
• If you run cAdvisor as a stand-alone DaemonSet for monitoring pods and containers, update it to v0.43.0 or later.
• If you deploy Java applications, prefer to use versions which fully support cgroup v2:
  • OpenJDK / HotSpot: jdk8u372, 11.0.16, 15 and later
  • IBM Semeru Runtimes: 8.0.382.0, 11.0.20.0, 17.0.8.0, and later
  • IBM Java: 8.0.8.6 and later
• If you are using the uber-go/automaxprocs package, make sure the version you use is v1.5.1 or higher.

Identify the cgroup version on Linux Nodes

The cgroup version depends on the Linux distribution being used and the default cgroup version configured on the OS. To check which cgroup version your distribution uses, run the stat -fc %T /sys/fs/cgroup/ command on the node:

stat -fc %T /sys/fs/cgroup/

For cgroup v2, the output is cgroup2fs.

For cgroup v1, the output is tmpfs.

What's next

• Learn more about cgroups
• Learn more about container runtime
• Learn more about cgroup drivers

7 - Container Runtime Interface (CRI)

The CRI is a plugin interface which enables the kubelet to use a wide variety of container runtimes, without having a need to recompile the cluster components.

You need a working container runtime on each Node in your cluster, so that the kubelet can launch Pods and their containers.

The Container Runtime Interface (CRI) is the main protocol for the communication between the kubelet and Container Runtime.

The Kubernetes Container Runtime Interface (CRI) defines the main gRPC protocol for the communication between the node components kubelet and container runtime.

The API

The kubelet acts as a client when connecting to the container runtime via gRPC. The runtime and image service endpoints have to be available in the container runtime, which can be configured separately within the kubelet by using the --image-service-endpoint command line flags.

For Kubernetes v1.30, the kubelet prefers to use CRI v1. If a container runtime does not support v1 of the CRI, then the kubelet tries to negotiate any older supported version. The v1.30 kubelet can also negotiate CRI v1alpha2, but this version is considered as deprecated. If the kubelet cannot negotiate a supported CRI version, the kubelet gives up and doesn't register as a node.

Upgrading

When upgrading Kubernetes, the kubelet tries to automatically select the latest CRI version on restart of the component. If that fails, then the fallback will take place as mentioned above. If a gRPC re-dial was required because the container runtime has been upgraded, then the container runtime must also support the initially selected version or the redial is expected to fail. This requires a restart of the kubelet.

What's next

• Learn more about the CRI protocol definition

8 - Garbage Collection

Garbage collection is a collective term for the various mechanisms Kubernetes uses to clean up cluster resources. This allows the clean up of resources like the following:

• Terminated pods
• Completed Jobs
• Objects without owner references
• Unused containers and container images
• Dynamically provisioned PersistentVolumes with a StorageClass reclaim policy of Delete
• Stale or expired CertificateSigningRequests (CSRs)
• Nodes deleted in the following scenarios:
  • On a cloud when the cluster uses a cloud controller manager
  • On-premises when the cluster uses an addon similar to a cloud controller manager
• Node Lease objects

Owners and dependents

Many objects in Kubernetes link to each other through owner references. Owner references tell the control plane which objects are dependent on others. Kubernetes uses owner references to give the control plane, and other API clients, the opportunity to clean up related resources before deleting an object. In most cases, Kubernetes manages owner references automatically.

Ownership is different from the labels and selectors mechanism that some resources also use. For example, consider a Service that creates EndpointSlice objects. The Service uses labels to allow the control plane to determine which EndpointSlice objects are used for that Service. In addition to the labels, each EndpointSlice that is managed on behalf of a Service has an owner reference. Owner references help different parts of Kubernetes avoid interfering with objects they don’t control.

Cascading deletion

Kubernetes checks for and deletes objects that no longer have owner references, like the pods left behind when you delete a ReplicaSet. When you delete an object, you can control whether Kubernetes deletes the object's dependents automatically, in a process called cascading deletion. There are two types of cascading deletion, as follows:

• Foreground cascading deletion
• Background cascading deletion

You can also control how and when garbage collection deletes resources that have owner references using Kubernetes finalizers.

Foreground cascading deletion

In foreground cascading deletion, the owner object you're deleting first enters a deletion in progress state. In this state, the following happens to the owner object:

• The Kubernetes API server sets the object's metadata.deletionTimestamp field to the time the object was marked for deletion.
• The Kubernetes API server also sets the metadata.finalizers field to foregroundDeletion.
• The object remains visible through the Kubernetes API until the deletion process is complete.

After the owner object enters the deletion in progress state, the controller deletes the dependents. After deleting all the dependent objects, the controller deletes the owner object. At this point, the object is no longer visible in the Kubernetes API.

During foreground cascading deletion, the only dependents that block owner deletion are those that have the ownerReference.blockOwnerDeletion=true field. See Use foreground cascading deletion to learn more.

Background cascading deletion

In background cascading deletion, the Kubernetes API server deletes the owner object immediately and the controller cleans up the dependent objects in the background. By default, Kubernetes uses background cascading deletion unless you manually use foreground deletion or choose to orphan the dependent objects.

See Use background cascading deletion to learn more.

Orphaned dependents

When Kubernetes deletes an owner object, the dependents left behind are called orphan objects. By default, Kubernetes deletes dependent objects. To learn how to override this behaviour, see Delete owner objects and orphan dependents.

Garbage collection of unused containers and images

The kubelet performs garbage collection on unused images every two minutes and on unused containers every minute. You should avoid using external garbage collection tools, as these can break the kubelet behavior and remove containers that should exist.

To configure options for unused container and image garbage collection, tune the kubelet using a configuration file and change the parameters related to garbage collection using the KubeletConfiguration resource type.

Container image lifecycle

Kubernetes manages the lifecycle of all images through its image manager, which is part of the kubelet, with the cooperation of cadvisor. The kubelet considers the following disk usage limits when making garbage collection decisions:

• HighThresholdPercent
• LowThresholdPercent

Disk usage above the configured HighThresholdPercent value triggers garbage collection, which deletes images in order based on the last time they were used, starting with the oldest first. The kubelet deletes images until disk usage reaches the LowThresholdPercent value.

Garbage collection for unused container images

As a beta feature, you can specify the maximum time a local image can be unused for, regardless of disk usage. This is a kubelet setting that you configure for each node.

To configure the setting, enable the ImageMaximumGCAge feature gate for the kubelet, and also set a value for the imageMaximumGCAge field in the kubelet configuration file.

The value is specified as a Kubernetes duration; Valid time units for the imageMaximumGCAge field in the kubelet configuration file are:

• "ns" for nanoseconds
• "us" or "µs" for microseconds
• "ms" for milliseconds
• "s" for seconds
• "m" for minutes
• "h" for hours

For example, you can set the configuration field to 12h45m, which means 12 hours and 45 minutes.

Container garbage collection

The kubelet garbage collects unused containers based on the following variables, which you can define:

• MinAge: the minimum age at which the kubelet can garbage collect a container. Disable by setting to 0.
• MaxPerPodContainer: the maximum number of dead containers each Pod can have. Disable by setting to less than 0.
• MaxContainers: the maximum number of dead containers the cluster can have. Disable by setting to less than 0.

In addition to these variables, the kubelet garbage collects unidentified and deleted containers, typically starting with the oldest first.

MaxPerPodContainer and MaxContainers may potentially conflict with each other in situations where retaining the maximum number of containers per Pod (MaxPerPodContainer) would go outside the allowable total of global dead containers (MaxContainers). In this situation, the kubelet adjusts MaxPerPodContainer to address the conflict. A worst-case scenario would be to downgrade MaxPerPodContainer to 1 and evict the oldest containers. Additionally, containers owned by pods that have been deleted are removed once they are older than MinAge.

Configuring garbage collection

You can tune garbage collection of resources by configuring options specific to the controllers managing those resources. The following pages show you how to configure garbage collection:

• Configuring cascading deletion of Kubernetes objects
• Configuring cleanup of finished Jobs

What's next

• Learn more about ownership of Kubernetes objects.
• Learn more about Kubernetes finalizers.
• Learn about the TTL controller that cleans up finished Jobs.

9 - Mixed Version Proxy

Kubernetes 1.30 includes an alpha feature that lets an API Server proxy a resource requests to other peer API servers. This is useful when there are multiple API servers running different versions of Kubernetes in one cluster (for example, during a long-lived rollout to a new release of Kubernetes).

This enables cluster administrators to configure highly available clusters that can be upgraded more safely, by directing resource requests (made during the upgrade) to the correct kube-apiserver. That proxying prevents users from seeing unexpected 404 Not Found errors that stem from the upgrade process.

This mechanism is called the Mixed Version Proxy.

Enabling the Mixed Version Proxy

Ensure that UnknownVersionInteroperabilityProxy feature gate is enabled when you start the API Server:

kube-apiserver \
--feature-gates=UnknownVersionInteroperabilityProxy=true \
# required command line arguments for this feature
--peer-ca-file=<path to kube-apiserver CA cert>
--proxy-client-cert-file=<path to aggregator proxy cert>,
--proxy-client-key-file=<path to aggregator proxy key>,
--requestheader-client-ca-file=<path to aggregator CA cert>,
# requestheader-allowed-names can be set to blank to allow any Common Name
--requestheader-allowed-names=<valid Common Names to verify proxy client cert against>,

# optional flags for this feature
--peer-advertise-ip=`IP of this kube-apiserver that should be used by peers to proxy requests`
--peer-advertise-port=`port of this kube-apiserver that should be used by peers to proxy requests`

# …and other flags as usual

Proxy transport and authentication between API servers

• The source kube-apiserver reuses the existing APIserver client authentication flags --proxy-client-cert-file and --proxy-client-key-file to present its identity that will be verified by its peer (the destination kube-apiserver). The destination API server verifies that peer connection based on the configuration you specify using the --requestheader-client-ca-file command line argument.

• To authenticate the destination server's serving certs, you must configure a certificate authority bundle by specifying the --peer-ca-file command line argument to the source API server.

Configuration for peer API server connectivity

To set the network location of a kube-apiserver that peers will use to proxy requests, use the --peer-advertise-ip and --peer-advertise-port command line arguments to kube-apiserver or specify these fields in the API server configuration file. If these flags are unspecified, peers will use the value from either --advertise-address or --bind-address command line argument to the kube-apiserver. If those too, are unset, the host's default interface is used.

Mixed version proxying

When you enable mixed version proxying, the aggregation layer loads a special filter that does the following:

• When a resource request reaches an API server that cannot serve that API (either because it is at a version pre-dating the introduction of the API or the API is turned off on the API server) the API server attempts to send the request to a peer API server that can serve the requested API. It does so by identifying API groups / versions / resources that the local server doesn't recognise, and tries to proxy those requests to a peer API server that is capable of handling the request.
• If the peer API server fails to respond, the source API server responds with 503 ("Service Unavailable") error.

How it works under the hood

When an API Server receives a resource request, it first checks which API servers can serve the requested resource. This check happens using the internal StorageVersion API.

• If the resource is known to the API server that received the request (for example, GET /api/v1/pods/some-pod), the request is handled locally.

• If there is no internal StorageVersion object found for the requested resource (for example, GET /my-api/v1/my-resource) and the configured APIService specifies proxying to an extension API server, that proxying happens following the usual flow for extension APIs.

• If a valid internal StorageVersion object is found for the requested resource (for example, GET /batch/v1/jobs) and the API server trying to handle the request (the handling API server) has the batch API disabled, then the handling API server fetches the peer API servers that do serve the relevant API group / version / resource (api/v1/batch in this case) using the information in the fetched StorageVersion object. The handling API server then proxies the request to one of the matching peer kube-apiservers that are aware of the requested resource.

  • If there is no peer known for that API group / version / resource, the handling API server passes the request to its own handler chain which should eventually return a 404 ("Not Found") response.

  • If the handling API server has identified and selected a peer API server, but that peer fails to respond (for reasons such as network connectivity issues, or a data race between the request being received and a controller registering the peer's info into the control plane), then the handling API server responds with a 503 ("Service Unavailable") error.