Kubernetes is a distributed cluster technology that manages a container-based system in a declarative manner using an API. Kubernetes is an open-source project governed by the Cloud Native Computing Foundation (CNCF) and counts all the largest cloud providers and software vendors as its contributors, along with a long list of community supporters.  

There are currently many learning resources to get started with the fundamentals of Kubernetes, but there is less information on how to manage Kubernetes infrastructure on an ongoing basis. This Refcard aims to deliver quickly accessible information for operators using any Kubernetes product.   

For an organization to deliver and manage Kubernetes clusters to every line of business and developer group, operations need to architect and manage both the core Kubernetes container orchestration, and the necessary auxiliary solutions (platform applications) — for example, monitoring, logging, and CI/CD pipeline. The CNCF maps out many of these solutions and groups them by category in the “CNCF Landscape.”   

The widespread adoption of cloud-native Kubernetes is reflected in these Gartner predictions: 

• “By 2025, Gartner estimates that over 95% of new digital workloads will be deployed on cloud-native platforms, up from 30% in 2021.” 

• “More than 85% of organizations will embrace a cloud-first principle by 2025 and will not be able to fully execute on their digital strategies without the use of cloud-native architectures and technologies.” 

Kubernetes differs from the orchestration offered by configuration management solutions in that it provides a declarative API that collects, stores, and reconciles desired state against observed state in an eventually consistent manner.  

A few traits of Kubernetes include: 

• Abstraction – Kubernetes abstracts the application orchestration from the infrastructure resource and as-a-service automation. This allows organizations to focus on the APIs of Kubernetes to manage an application at scale in a highly available manner instead of the underlying infrastructure resources. 
• Declarative – Kubernetes’ control plane decides how the hosted application is deployed and scaled on the underlying fabric. A user simply defines the logical configuration of the Kubernetes object, and the control plane takes care of the implementation. 
• Immutable – Different versions of services running on Kubernetes are completely new and not swapped out. Objects in Kubernetes, say different versions of a pod, are changed by creating new objects.  

Figure 1: Simplified Kubernetes architecture and components 

Control Plane, Nodes, and Persistent Storage 

Kubernetes’ basic architecture requires a number of components. 

API Server 

The Kubernetes API server handles all requests coming into the cluster. Users, as well as other Kubernetes components, send events to the Kubernetes API Server through HTTPS (port 443). The API Server then processes events and updates the Kubernetes persistent data store, usually etcd. The API Server also performs authentication and authorization depending on how the cluster was configured.  

https://kubernetes.io/docs/concepts/architecture/master-node-communication/ 

Control Plane Datastore

All Kubernetes cluster data is stored in a single datastore. The default datastore is etcd. . A best practice is to ensure that there are multiple etcd instances to ensure high availability of the cluster. A loss of etcd storage will result in a loss of the cluster's state, so etcd should be backed up for disaster recovery. Since etcd is the single source of truth for the cluster, it's imperative to secure it against malicious actors. 

https://kubernetes.io/docs/tasks/administer-cluster/configure-upgrade-etcd/ 

Controller Manager 

Kubernetes uses a control plane with several types of controllers to perform non-terminating control loops that observe the state of Kubernetes objects and reconcile it with the desired state. This includes a wide variety of functions such as invoking pod admission controllers, setting pod defaults, or injecting sidecar containers into pods, all according to the configuration of the cluster. 

https://kubernetes.io/docs/concepts/architecture/controller/ 

Kubelet (Agent) 

Each node in a Kubernetes cluster has an agent called the Kubelet. The Kubelet manages the container runtime (e.g., containerdt) on individual nodes. The kubelet reads workload definitions from the API server, ensures the requested workloads run and stay healthy, and reports the status of both workloads and the node itself back to the API server. 

https://kubernetes.io/docs/concepts/overview/components/

Scheduler 

Kubernetes relies on a sophisticated algorithm to schedule Kubernetes objects. The scheduling takes into account attributes from a given workload (e.g., resource requirements) and applies custom prioritization settings (e.g., affinity rules) to provision pods on particular Kubernetes worker nodes. 

https://kubernetes.io/docs/concepts/scheduling-eviction/kube-scheduler/ 

Kube-Proxy 

Each Kubernetes cluster worker node has a network proxy for connectivity to application services in the cluster. The proxy reflects the services defined in the cluster and enables simple stream and round-robin forwarding to a set of backends, regardless of which node they reside in the cluster. 

https://kubernetes.io/docs/concepts/overview/components/#kube-proxy 

Other Kubernetes Components (Platform Applications) 

To run the most basic Kubernetes cluster, a number of additional components and platform applications are required. 

DNS 

Production deployments of Kubernetes include a DNS server, which can be used for service discovery. All services and pods in a cluster are assigned a domain name that is resolved by the DNS. This DNS is ordinarily handled by Kubernetes DNS, which by default is backed by popular services such as CoreDNS (coredns.io). With Kubernetes DNS configured, Services (and pods) can be addressed by a defined naming convention in the form of A/AAAA or SRV records for discernable access by clients or other Services in the cluster. 

kubectl 

The official command line for Kubernetes is called kubectl. All industry-standard Kubernetes commands start with kubectl.  

Metrics Server and Metrics API 

Kubernetes has two resources for giving usage metrics to users and tools. First, Kubernetes can include the metrics server, which is the centralized aggregation point for Kubernetes metrics. Second is the Kubernetes metrics API, which provides an API to access these aggregated metrics. 

Web UI (Dashboard)  

Kubernetes has an official GUI called Dashboard. That GUI is distinct from vendor GUIs that have been designed for specific Kubernetes derivative products. Note that the Kubernetes Dashboard release version does not necessarily match the Kubernetes release version. 

Kubernetes has a number of constructs for defining and managing objects on the cluster: 

Kubernetes has a number of points to extend its core functionality:  

Below are some commands useful for IT professionals getting started with Kubernetes. A full list of Kubectl commands: https://kubernetes.io/docs/reference/generated/kubectl/kubectl-commands 

Making Life Easier 

Finding Kubernetes command short name: 

$ kubectl describe 

Find out more: 

• Using Kubectl Aliases – https://github.com/ahmetb/kubectl-aliases 
• Switching among Kubernetes clusters and namespaces (context switching) – https://github.com/ahmetb/kubectx 
• Tailing logs from multiple pods – https://github.com/wercker/stern 

Setting a Bash prompt to mirror Kubernetes context:  

prompt_set() { 
if [ "$KUBECONFIG" != "" ]; then  
  PROMPT_KUBECONTEXT="k8s:$(kubectl config current-context 2>/dev/null)\n" 
fi 
PS1="${PROMPT_KUBECONTEXT}[\u@\h \W]\$ " 
} 
k8s:admin@local 
[kensey@sleeper-service ~]$ 

Commands

Sometimes it's necessary to perform maintenance on underlying nodes. In those cases, it's important to use eviction and ensure the application owners have set a pod disruption budget. 

To properly evict running pods from a node that requires maintenance, use: 

1. Cordon node: $ kubectl cordon $NODENAME 
2. Drain node: $ kubectl drain $NODENAME --ignore-daemonsets 
   1. Automatically cordons the node (removes it from the possible schedulable node pool) 
   2. Respects PodDisruptionBudgets 
   3. Will not error out for running DaemonSet pods 
3. Uncordon node: $ kubectl uncordon $NODENAME 

It's important to devise a test plan and a run book to ensure any new clusters meet operational guidelines. This enables standard operating procedures (SOPs) that address needs such as consistency or compliance, facilitating better operational metrics for uptime and availability. 

Below are samples from D2iQ's Kubernetes test plan, which is used when validating new clusters, Kubernetes versions, platform applications, or workloads running on the cluster. After each step, validate that the cluster is working properly.  

Test Clusters 

1. Provision a highly available Kubernetes cluster with 100 Kubernetes nodes 
2. Scale that cluster down to 30 nodes after it has finished provisioning 
3. Provision three additional highly available Kubernetes clusters with five nodes 
4. Scale all three clusters (in parallel) to 30 nodes simultaneously 
5. Provision 16 more highly available Kubernetes clusters with 30 nodes 
6. Kill five Kubernetes nodes on a single cluster simultaneously 
7. Kill three control-plane nodes on a single cluster (fewer if the nodes will not automatically reprovision) 
8. Kill the etcd leader 

Test Workloads 

1. Provision storage (potentially using CSI and a specific driver) 
2. Test different provider-specific storage features 
3. Run the e2e cluster loader with the higher end of pods per node (35-50) 
4. Run Kubernetes conformance testing to stress the cluster 
5. Provision Helm to the Kubernetes cluster 
6. Test specific Kubernetes workloads (using Helm charts) 
7. Deploy NGINX 
8. Deploy Redis 
9. Deploy Postgres 
10. Deploy RabbitMQ 
11. Deploy services with type=loadbalancer to test load balancer automation 
12. Test different provider-specific load balancer features 
13. Expose a service using Ingress  

Troubleshoot With Kubectl 

The Kubernetes command line kubectl provides many of the resources needed to debug clusters. 

Check Permissions 

Pending and Waiting Pods 

Pending —> Check Resources 

Remove pods from the environment. 

Waiting —> Incorrect Image Information 

spec:    
  containers:   
  — name: example    
    image: url:port/image:v 

docker pull <image> 

Also check that the secret information is correct. 

Crash Looping Deployment 

• Roll back deployment 
• Troubleshoot 

General Debugging 

Watch the stdout of a container: 

$ kubectl logf -f bash-rand-77d55b86c7-bntxs 

Launch a temporary pod with an interactive shell: 

$ kubectl run busybox --rm -i --tty --restart=Never --image busybox -- /bin/sh 

Copy files into and out of containers (note that this requires a tar binary in the container): 

$ kubectl cp default/bash-rand-77d55b86c7-bntxs:/root/rand . 
tar: Removing leading `/' from member names 
 
$ cat rand  
567bea045d8b80cd6d007ced02849ac4 

Increase the verbosity of kubectl output (99 is the most verbose — “get everything”): 

$ kubectl -v 99 get nodes 
I1211 11:10:28.611959   24842 loader.go:359] Config loaded from file /home/kensey/bootkube/cluster/auth/kubeconfig 
I1211 11:10:28.612482   24842 loader.go:359] Config loaded from file /home/kensey/bootkube/cluster/auth/kubeconfig 
I1211 11:10:28.614383   24842 loader.go:359] Config loaded from file /home/kensey/bootkube/cluster/auth/kubeconfig 
I1211 11:10:28.617867   24842 loader.go:359] Config loaded from file /home/kensey/bootkube/cluster/auth/kubeconfig 
I1211 11:10:28.629567   24842 round_trippers.go:405] GET https://192.168.122.138:6443/api/v1/nodes?limit=500 200 OK in 11 milliseconds 
I1211 11:10:28.630279   24842 get.go:558] no kind is registered for the type v1beta1.Table in scheme "k8s.io/kubernetes/pkg/api/legacyscheme/scheme.go:29" 
NAME  STATUS   ROLES   AGE   VERSION 

Get more information about Kubernetes resources: 

$ kubectl explain crd 
KIND:     CustomResourceDefinition 
VERSION:  apiextensions.k8s.io/v1beta1 

DESCRIPTION: 
     CustomResourceDefinition represents a resource that should be exposed on 
     the API server. Its name MUST be in the format <.spec.name>.<.spec.group>. 
FIELDS: 
[...] 
Control output: 
$ kubectl get deploy bash-rand -o yaml 
apiVersion: extensions/v1beta1 
kind: Deployment 
metadata: 
  annotations: 
    deployment.kubernetes.io/revision: "1" 
  creationTimestamp: 2018-12-11T05:20:32Z 
  generation: 1 
  labels: 
    run: bash-rand 
[...] 

Troubleshoot With jq 

Kubectl provides an option to output results in JSON format, which allows for additional parsing and filtering. This is beneficial when writing scripts or monitoring the state of resources in a Kubernetes cluster. A popular open-source utility called jq makes it easy to parse JSON from the command line. Instead of writing code to parse the JSON objects, the output can be piped into the jq utility to filter out resources that meet certain criteria. For example, kubectl can be used to print out all pods in json format, but jq adds value by parsing out only pods with a specific start date or container image. 

Basic usage: 

[some JSON content] | jq [flags] [filter expression] 

Use the kubectl json format and pipe to the jq utility using a simple dot filter to pretty-print the output: 

$ kubectl get nodes -o json | jq '.' 

Use the hierarchy to filter the input so only node labels are displayed: 

$ kubectl get nodes -o json | jq '.items[].metadata.labels' 

Advanced Functions 

Find pods with containers that have not passed a readiness probe: 

• select – cherry-pick a piece of the input by criteria and use a boolean expression to match desired value 
• Pipe the output of one filter into another filter to clean up the results 

$ kubectl get pods -o json | jq '.items[] | select( .status.containerStatuses[].ready == false ) | .metadata.name' 

Troubleshoot With Curl 

Adding headers to requests are often used for: 

• Setting accepted content types for replies. 
• Setting content type of posted content. 
• Injecting bearer auth tokens. 

 $ curl --header "Authorization: Bearer [token]" [API server URL] 

 Build and submit requests: 

• GET – Request is contained in the URL itself (default method) and used to read/list/watch resources 
• POST – Submit a data blob to create resources 
• PATCH – Submit a data blob to merge-update resources 
• PUT – Submit a data blob to replace a resource 
• DELETE – Submit options to control deletion of a resource 

$ curl --cert client.cert --key client.key --cacert cluster-ca.cert \ 
 https://192.168.100.10:6443/api/v1/namespaces/default/pods 
  { 
   "kind": "PodList", 
   "apiVersion": "v1", 
   "metadata": { 
   "selfLink": "/api/v1/namespaces/default/pods", 
   "resourceVersion": "126540" 
    }, 
  "items": [ 

bash while read loop: 

$ while read -r serverevent; do echo "$serverevent" \ 
| jq '.["operation"] = .type | .["node"] = .object.metadata.name | { "operation": .operation, "node": .node}' ; \ 
done < <(curl -sN --cert client.cert --key client.key --cacert cluster-ca.cert \ 
https://192.168.100.10:6443/api/v1/nodes?watch=true) 
{ 
  "operation": "ADDED", 
  "node": "192.168.100.10" 
} 
[...] 
 
bash coproc: 
#!/bin/bash 
 
coproc curl -sN --cacert cluster-ca.cert --cert ./client.cert --key ./client.key \ 
https://192.168.100.10:6443/api/v1/nodes?watch=true 
 
exec 5<&${COPROC[0]} 
 
while read -ru 5 serverevent; do 
  if [[ $(echo $serverevent | jq -r '.type') == "ADDED" ]]; then 
    echo "Added node $(echo $serverevent | jq -r '.object.metadata.name') in namespace \ 
$(echo $serverevent | jq '.object.metadata.namespace')" 
  fi 
done 
 
trap 'kill -TERM $COPROC_PID' TERM INT 

A production Kubernetes cluster will typically have three etcd nodes, which provides high availability and can withstand the loss of a single etcd node. The etcdctl utility provides a number of commands for running operations on the etcd cluster: 

Security is fundamental for Kubernetes day-two operations. Users should maintain basic security for the clusters. Here is a security checklist for Kubernetes administrators:  

• Kubernetes patch releases — i.e., 1.x.y (ensure it has all CVE fixes) 
  • A new version of Kubernetes is released every three months. Patches come out at regular intervals for the latest three versions and have bug fixes. 
  • Versions must be kept up to date due to the lack of widespread support for older versions of Kubernetes and the bug fixes in each patch release. Note: Kubernetes constantly releases security fixes, and unpatched clusters are vulnerable.  
• Transport Layer Security (TLS) 
  • Kubernetes components must use an encrypted connection. Make sure that Kubernetes clusters have end-to-end TLS enabled. 
• Kubernetes RBAC and authentication  
  • Kubernetes has several forms of access management, including role-based access control (RBAC). Make sure that RBAC is enabled and that users are assigned proper roles. 
• Network policies enabled 
  • A Kubernetes network is flat. Every pod is given an IP address that can communicate with all other pods on its namespace. It is possible to use network policies to limit the interactions among pods and lock down this cross-communication. 
• Different clusters or namespaces based on the security profile 
  • Namespaces can create logical clusters in a single physical cluster, but they only provide soft isolation. Do not include vastly different services in the same cluster. 
  • Using different Kubernetes clusters reduces the potential for vulnerabilities to affect all systems. Also, large clusters with unrelated services become harder to upgrade, which violates the first item in this checklist.  
• Implement resource quotas 
  • To prevent “noisy neighbors” and potential denial of service situations, do not let containers run without an upper bound on resources. By default, all resources are created with unbounded CPU and memory limits. To avoid the default behavior, assign resource quotas at the namespace level, which are critical in preventing overconsumption of resources. 
• Limit access to insecure API server ports 
  • The API Server is the main means to communicate with Kubernetes components. Besides enabling authentication and RBAC, lock down all unsecure API server ports.  
• Limit access of pod creation for users 
  • One of the most widely used vectors for attacking container management systems are the containers themselves. To ensure the containers are secure and protect the system from attacks:  
    • Limit who can create pods. 
    • Limit the use of unknown or unmaintained libraries.  
    • Use a private container registry and tagged container images, keeping tagged images immutable.  
• Secure Dashboard 
  • The Kubernetes Dashboard is a functional utility for users, but it can also be a vector for malicious attacks — limit its exposure as much as possible. 
  • Employ three specific tactics: 
    • Do not expose the Dashboard to the Internet 
    • Ensure the Dashboard ServiceAccount is not open and accessible to users 
    • Configure the login page and enable RBAC 

• Tracking Kubernetes Releases: https://github.com/kubernetes/sig-release/tree/master/releases  
• Enhancements Tracking Sheet 
• Bug Triage Tracking Sheet 
• Release Calendar 
• Tracking Kubernetes Enhancement Proposals: https://github.com/kubernetes/enhancements/tree/master/keps 
• Check if the Kubernetes distribution or installer has been through conformance testing: https://docs.google.com/spreadsheets/d/1LxSqBzjOxfGx3cmtZ4EbB_BGCxT_wlxW_xgHVVa23es/edit#gid=0 
• Run end-to-end tests to validate a Kubernetes cluster: https://github.com/kubernetes/test-infra 
• KUTTL – Kubernetes declarative test tool: https://github.com/kudobuilder/kuttl 
• CIS Security Benchmark tool: https://github.com/mesosphere/kubernetes-security-benchmark 
• YAML Templates: https://cheatsheet.dennyzhang.com/kubernetes-yaml-templates 
• Use jq to process JSON output from the command line: https://stedolan.github.io/jq/