clusterlint/checks.md

21 KiB

Clusterlint flags issues with workloads deployed in a cluster. These issues might cause workload downtime during maintenance or upgrade to a new Kubernetes version, and could complicate the maintenance or upgrade itself. DigitalOcean runs a cluster linter check before a cluster upgrade. You can also run these checks yourself from the Operational Readiness Check section of a cluster's Overview node.

Default Namespace

  • Name: default-namespace
  • Groups: basic

Namespaces are a way to limit the scope of the resources that subsets of users within a team can create. While a default namespace is created for every Kubernetes cluster, we do not recommend adding all created resources into the default namespace because of the risk of privilege escalation, resource name collisions, latency in operations as resources scale up, and mismanagement of Kubernetes objects. Having namespaces lets you enable resource quotas can be enabled to track node, CPU and memory usage for individual teams.

Example

# Not recommended: Defining resources with no namespace, which adds them to the default.
apiVersion: v1
kind: Pod
metadata:
  name: mypod
  labels:
    name: mypod
spec:
  containers:
  - name: mypod
    image: nginx:1.17.0

How to Fix

# Recommended: Explicitly specify a namespace in the object config
apiVersion: v1
kind: Pod
metadata:
  name: mypod
  namespace: test
  labels:
    name: mypod
spec:
  containers:
  - name: mypod
    image: nginx:1.17.0

Latest Tag

  • Name: latest-tag
  • Groups: basic

We do not recommend using container images with the latest tag or not specifying a tag in the image (which defaults to latest), as this leads to confusion around the version of image used. Pods get rescheduled often as conditions inside a cluster change, and upon a reschedule, you may find that the images' versions have changed to use the latest release, which can break the application and make it difficult to debug errors. Instead, update segments of the application individually using images pinned to specific versions.

Example

# Not recommended: Not specifying an image tag, or using "latest"
spec:
  containers:
  - name: mypod
    image: nginx
  - name: redis
    image: redis:latest

How to Fix

# Recommended: Explicitly specify a tag or digest
spec:
  containers:
  - name: mypod
    image: nginx:1.17.0
  - name: redis
    image: redis@sha256:dca057ffa2337682333a3aba69cc0e7809819b3cd7fc78f3741d9de8c2a4f08b

CronJob Concurrency

  • Name: cronjob-concurrency
  • Groups: basic

We do not recommend having a concurrencyPolicy of Allow for CronJob resources. If a CronJob-managed Pod does not execute to completion within the expected window, it is possible that multiple Pods pile up over time, leading to several Pods stuck in a pending state and possible resource contention. Instead, prefer Forbid, which skips execution of a new job if the previous job has not exited, or Replace, which replaces the still-running job with a new job if it has not yet exited.

Example

# Not recommended: Having a concurrency policy of Allow
apiVersion: batch/v1beta1
kind: CronJob
metadata:
   name: mycron
spec:
  concurrencyPolicy: Allow

How to Fix

# Recommended: Having a concurrency policy of Forbid or Replace
apiVersion: batch/v1beta1
kind: CronJob
metadata:
   name: mycron
spec:
  concurrencyPolicy: Replace

Privileged Containers

  • Name: privileged-containers
  • Groups: security

Use the privileged mode for trusted containers only. Because the privileged mode allows container processes to access the host, malicious containers can extensively damage the host and bring down services on the cluster. If you need to run containers in privileged mode, test the container before using it in production. For more information about the risks of running containers in privileged mode, please refer to the Kubernetes security context documentation.

Example

# Not recommended: Using privileged mode instead of granting capabilities when it's not necessary
spec:
  containers:
  - name: mypod
    image: nginx
    securityContext:
      privileged: true

How to Fix

# Recommended: Explicitly add only the needed capabilities to the container
spec:
  containers:
  - name: mypod
    image: nginx
    securityContext:
      capabilities:
        add:
        - NET_ADMIN

Run As Non-Root

  • Name: run-as-non-root
  • Groups: security

If containers within a pod are allowed to run with the process ID (PID) 0, then the host can be subjected to malicious activity. We recommend using a user identifier (UID) other than 0 in your container image for running applications. You can also enforce this in the Kubernetes pod configuration as shown below.

Example

# Not recommended: Doing nothing to prevent containers from running under UID 0
spec:
  containers:
  - name: mypod
    image: nginx

How to Fix

# Recommended: Ensure containers do not run as root
spec:
  securityContext:
    runAsNonRoot: true
  containers:
  - name: mypod
    image: nginx

Fully Qualified Image

  • Name: fully-qualified-image
  • Groups: basic

Docker is the most popular runtime for Kubernetes. However, Kubernetes supports other container runtimes as well, such as containerd and CRI-O. If the registry is not prepended to the image name, docker assumes docker.io and pulls it from Docker Hub. However, the other runtimes will result in errors while pulling images. To maintain portability, we recommend using a fully qualified image name. If the underlying runtime is changed and the object configs are deployed to a new cluster, having fully qualified image names ensures that the applications do not break.

Example

# Not recommended: Failing to specify the registry in the image name
spec:
  containers:
  - name: mypod
    image: nginx:1.17.0

How to Fix

# Recommended: Provide the registry name in the image
spec:
  containers:
  - name: mypod
    image: docker.io/nginx:1.17.0

Node Name Selector

  • Name: node-name-pod-selector
  • Groups: doks

On upgrade of a cluster on DOKS, the worker nodes' hostname changes. So, if your pod spec relies on the hostname to schedule pods on specific nodes, pod scheduling will fail after the upgrade.

Example

# Not recommended: Using a raw DigitalOcean resource name in the nodeSelector
apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    env: test
spec:
  containers:
  - name: nginx
    image: nginx
  nodeSelector:
    kubernetes.io/hostname: pool-y25ag12r1-xxxx

How to Fix

# Recommended: Use the DOKS-specific node pool label
apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    env: test
spec:
  containers:
  - name: nginx
    image: nginx
  nodeSelector:
    doks.digitalocean.com/node-pool: pool-y25ag12r1

Admission Controller Webhook

  • Name: admission-controller-webhook
  • Groups: basic

Admission control webhooks can disrupt normal cluster operations. Specifically, this happens when an admission control webhook targets a service that:

  • Does not exist
  • Is in a namespace that does not exist

Example

# Error: Configure a webhook pointing at a service that does not exist
apiVersion: admissionregistration.k8s.io/v1beta1
kind: ValidatingWebhookConfiguration
metadata:
  name: sample-webhook.example.com
webhooks:
- name: sample-webhook.example.com
  rules:
  - apiGroups:
    - ""
    apiVersions:
    - v1
    operations:
    - CREATE
    resources:
    - pods
    scope: "Namespaced"
  clientConfig:
    service:
      namespace: webhook
      name: missing-webhook-server
      path: /pods
  admissionReviewVersions:
  - v1beta1
  timeoutSeconds: 1
  failurePolicy: Fail

How to Fix

Point the webhook at the correct service.

apiVersion: admissionregistration.k8s.io/v1beta1
kind: ValidatingWebhookConfiguration
metadata:
  name: sample-webhook.example.com
webhooks:
- name: sample-webhook.example.com
  rules:
  - apiGroups:
    - ""
    apiVersions:
    - v1
    operations:
    - CREATE
    resources:
    - pods
    scope: "Namespaced"
  clientConfig:
    service:
      namespace: webhook
      name: webhook-server
      path: /pods
  admissionReviewVersions:
  - v1beta1
  timeoutSeconds: 1
  failurePolicy: Fail
  namespaceSelector:
    matchExpressions:
      - key: "skip-webhooks"
        operator: "DoesNotExist"

Admission Controller Webhook Replacement

  • Name: admission-controller-webhook-replacement
  • Groups: doks

Admission control webhooks can disrupt upgrade and node replacement operations by preventing system components from starting. Specifically, this happens when an admission control webhook:

  • Has failurePolicy set to Fail
  • Targets a service other than the Kubernetes apiserver
  • Applies to both kube-system and the namespace of the targeted service
  • Has rules applicable to v1, apps/v1, apps/v1beta1 or apps/v1beta2 resources

Example

# Not recommended: Configure a webhook with a failurePolicy set to "Fail"
apiVersion: admissionregistration.k8s.io/v1beta1
kind: ValidatingWebhookConfiguration
metadata:
  name: sample-webhook.example.com
webhooks:
- name: sample-webhook.example.com
  rules:
  - apiGroups:
    - "*"
    apiVersions:
    - "*"
    operations:
    - CREATE
    resources:
    - pods
    scope: "Namespaced"
  clientConfig:
    service:
      namespace: webhook
      name: webhook-server
      path: /pods
  admissionReviewVersions:
  - v1beta1
  timeoutSeconds: 1
  failurePolicy: Fail

How to Fix

Use one of the following options:

  • Use the Ignore failurePolicy.
  • Use an apiserver extension as your webhook service.
  • Explicitly exclude the kube-system namespace.
  • Explicitly exclude the webhook service's namespace.
  • Explicitly include the resource api group and version in the rules. If you have configured webhooks for CRDs, we recommend that you explicitly specify the rules instead of generally applying them to all resources.
# Recommended: Exclude objects in the `webhook` namespace by explicitly specifying a namespaceSelector.

apiVersion: v1
kind: Namespace
metadata:
  name: webhook
  labels:
    skip-webhooks: "yes"

---
apiVersion: admissionregistration.k8s.io/v1beta1
kind: ValidatingWebhookConfiguration
metadata:
  name: sample-webhook.example.com
webhooks:
- name: sample-webhook.example.com
  rules:
  - apiGroups:
    - ""
    apiVersions:
    - v1
    operations:
    - CREATE
    resources:
    - pods
    scope: "Namespaced"
  clientConfig:
    service:
      namespace: webhook
      name: webhook-server
      path: /pods
  admissionReviewVersions:
  - v1beta1
  timeoutSeconds: 1
  failurePolicy: Fail
  namespaceSelector:
    matchExpressions:
      - key: "skip-webhooks"
        operator: "DoesNotExist"

Admission Controller Webhook Timeout

  • Name: admission-controller-webhook-timeout
  • Groups: doks

Admission control webhook timeouts can block upgrades, when the API call times out, due to an incorrectly configured TimeoutSeconds value. Since webhooks inherently add to API latency, we must stay within the recommended range in order for API requests to be successful. Specifically, this happens when an admission control webhook does not respond within 29 seconds.

Example

# Error: Configure a webhook with a TimeoutSeconds value greater than 29 seconds.
apiVersion: admissionregistration.k8s.io/v1beta1
kind: ValidatingWebhookConfiguration
metadata:
  name: sample-webhook.example.com
webhooks:
- name: sample-webhook.example.com
  rules:
  - apiGroups:
    - ""
    apiVersions:
    - v1
    operations:
    - CREATE
    resources:
    - pods
    scope: "Namespaced"
  clientConfig:
    service:
      namespace: webhook
      name: webhook-server
      path: /pods
  admissionReviewVersions:
  - v1beta1
  timeoutSeconds: 60

How to Fix

Set the TimeoutSeconds value to anything within the 1 to 29 second range.

apiVersion: admissionregistration.k8s.io/v1beta1
kind: ValidatingWebhookConfiguration
metadata:
  name: sample-webhook.example.com
webhooks:
- name: sample-webhook.example.com
  rules:
  - apiGroups:
    - ""
    apiVersions:
    - v1
    operations:
    - CREATE
    resources:
    - pods
    scope: "Namespaced"
  clientConfig:
    service:
      namespace: webhook
      name: webhook-server
      path: /pods
  admissionReviewVersions:
  - v1beta1
  timeoutSeconds: 10

DOBS Pod Owner

  • Name: dobs-pod-owner
  • Groups: doks

This check ensures that any pod that references a DigitalOcean Block Storage Volume is owned by a StatefulSet. We want such pods to be owned by a StatefulSet because:

  • The Eviction API does not respect deployment strategies. It only cares about pod disruption budgets (PDBs). So, if you do not set it right, you can end up with multiple volume-using pods running concurrently. This can lead to stuck deployments if they happen to come up on different nodes in the best case, and data corruption if they come up on the same node and end up writing to the same volume concurrently. For more context, see Disruptions in the Kubernetes documentation.

  • Manual deletes do not care about PDBs at all. So, all pods from a Deployment, for instance, are deleted and brought up at the same time. A StatefulSet, on the other hand, always ensures “at most” guarantees.

Example

# Not recommended: Pods that refer to DOBS volumes should be owned by a StatefulSet
apiVersion: v1
kind: Pod
metadata:
  name: mypod
  namespace: test
  labels:
    name: mypod
spec:
  containers:
  - name: mypod
    image: nginx:1.17.0
    volumeMounts:
    - mountPath: "/data"
      name: all-data
  volumes:
    - name: my-dobs-volume
      persistentVolumeClaim:
        claim-name: do-csi-pvc

How to Fix

# Recommended: Ensure that a StatefulSet is used to configure pods referencing DOBS volumes
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: web
  namespace: test
spec:
  selector:
    matchLabels:
      app: mypod
  serviceName: "nginx"
  replicas: 1
  template:
    metadata:
      labels:
        app: mypod
    spec:
      containers:
      - name: nginx
        image: nginx:1.17.0
        volumeMounts:
        - name: all-data
          mountPath: "/data"
  volumeClaimTemplates:
  - metadata:
      name: all-data
    spec:
      accessModes: [ "ReadWriteOnce" ]
      resources:
        requests:
          storage: 1Gi
      storageClassName: do-block-storage

Pod State

  • Name: pod-state
  • Groups: workload-health

This checks for unhealthy pods in a cluster. This check is not run by default. Specify a group name or a check name to run this check.

HostPath Volume

  • Name: hostpath-volume
  • Groups: basic

Using hostPath volumes is best avoided because:

  • Pods with an identical configuration (such as those created from a podTemplate) intended to behave identically to one another regardless of their deployment will in fact behave differently from node to node due to differences in the files present on the nodes themselves.
  • Resource-aware scheduling is not be able to account for resources used by a hostPath volume.
  • The files created on the hosts are only writable by root; you will need to run your process as root in a privileged container or modify the file permissions on the host to be able to write to a hostPath volume.

For more details about hostPath volumes, please refer to the Kubernetes documentation

Example

# Not recommended: Using a hostPath volume
apiVersion: v1
kind: Pod
metadata:
  name: test-pd
spec:
  containers:
  - image: docker.io/nginx:1.17.0
    name: test-container
    volumeMounts:
    - mountPath: /test-pd
      name: test-volume
  volumes:
  - name: test-volume
    hostPath:
      path: /data
      type: Directory

How to Fix

# Recommended: Use other volume sources. See https://kubernetes.io/docs/concepts/storage/volumes/
apiVersion: v1
kind: Pod
metadata:
  name: test-pd
spec:
  containers:
  - image: docker.io/nginx:1.17.0
    name: test-container
    volumeMounts:
    - mountPath: /test-pd
      name: test-volume
  volumes:
  - name: test-volume
    cephfs:
      monitors:
        - 10.16.154.78:6789
      user: admin
      secretFile: "/etc/ceph/admin.secret"
      readOnly: true

Unused Persistent Volume

  • Name: unused-pv
  • Groups: basic

This check reports all the persistent volumes in the cluster that are not claimed by a PersistentVolumeClaim (PVC) in any namespace. You can clean up the cluster based on this information and there will be fewer objects to manage.

How to Fix

kubectl delete pv <unused pv>

Unused Persistent Volume Claims

  • Name: unused-pvc
  • Groups: basic

This check reports all the PVCs in the cluster that are not referenced by pods in the respective namespaces. You can clean up the cluster based on this information.

How to Fix

kubectl delete pvc <unused pvc>

Unused Config Maps

  • Name: unused-config-map
  • Groups: basic

This check reports all the config maps in the cluster that are not referenced by pods in the respective namespaces. You can clean up the cluster based on this information.

How to Fix

kubectl delete configmap <unused config map>

Unused Secrets

  • Name: unused-secret
  • Groups: basic

This check reports all the secret names in the cluster that are not referenced by pods in the respective namespaces. You can clean up the cluster based on this information.

How to Fix

kubectl delete secret <unused secret name>

Resource Requirements

  • Name: resource-requirements
  • Groups: basic

When you specify resource limits for containers, the scheduler can make better decisions about which nodes to place pods on, and handle contention for resources on a node in a specified manner.

Example

# Not recommended: Scheduling pods without specifying any resource limits
apiVersion: v1
kind: Pod
metadata:
  name: test
spec:
  containers:
  - image: docker.io/nginx:1.17.0
    name: test-container

How to Fix

# Recommended: Specify resource requests and limits
apiVersion: v1
kind: Pod
metadata:
  name: test-pd
spec:
  containers:
  - image: docker.io/nginx:1.17.0
    name: test-container
    resources:
      limits:
        cpu: 102m
      requests:
        cpu: 102m

Bare Pods

  • Name: bare-pods
  • Groups: basic

When the node that a pod is running on reboots or fails, the pod is terminated and will not be restarted. However, a job will create new pods to replace terminated ones. For this reason, we recommend that you use a job, deployment, or StatefulSet rather than a bare pod, even if your application requires only a single pod.

Example

# Not recommended: Deploying a bare pod without any deployment parameters
apiVersion: v1
kind: Pod
metadata:
  name: mypod
  namespace: test
  labels:
    name: mypod
spec:
  containers:
  - name: mypod
    image: nginx:1.17.0

How to Fix

# Recommended: Configure pods as part of a deployment, job, or StatefulSet
apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-deployment
  namespace: test
  labels:
    app: nginx
spec:
  replicas: 1
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
      - name: nginx
        image: nginx:1.7.9

Node Labels and Taints

  • Name: node-labels-and-taints
  • Groups: doks

When you upgrade a DOKS cluster, all worker nodes are replaced, and replacement nodes do not retain any custom labels or taints you previously set on the nodes. This check reports any labels or taints that will be lost on upgrade.

DOKS provides persistent node pool labels. Adding a custom label to a node pool ensures that the label is propagated to the worker nodes in the node pool after replacement or upgrade.

How to Fix

kubectl label node <node-name> <label-key>-
kubectl taint node <node-name> <taint-key>-

Note the trailing - on the key; this causes kubectl to delete the label or taint.

Images From GitHub Packages Docker Registry

  • Name: docker-pkg-github-com-registry
  • Groups: containerd, doks

containerd cannot pull container images from docker.pkg.github.com due to a protocol mismatch. DOKS 1.20 and newer use containerd as the container runtime, and are unable to use images from docker.pkg.github.com.

Example

# Not supported: Using "docker.pkg.github.com" as container registry
spec:
  containers:
  - name: redis
    image: docker.pkg.github.com/redis/redis/redis:6

How to Fix

GitHub recommends migrating to GitHub Container Registry, which is compatible with containerd.