Policies

• 1: Limit Ranges
• 2: Resource Quotas
• 3: Process ID Limits And Reservations
• 4: Node Resource Managers

1 - Limit Ranges

By default, containers run with unbounded compute resources on a Kubernetes cluster. With resource quotas, cluster administrators can restrict resource consumption and creation on a namespace basis. Within a namespace, a Pod or Container can consume as much CPU and memory as defined by the namespace's resource quota. There is a concern that one Pod or Container could monopolize all available resources. A LimitRange is a policy to constrain resource allocations (to Pods or Containers) in a namespace.

A LimitRange provides constraints that can:

• Enforce minimum and maximum compute resources usage per Pod or Container in a namespace.
• Enforce minimum and maximum storage request per PersistentVolumeClaim in a namespace.
• Enforce a ratio between request and limit for a resource in a namespace.
• Set default request/limit for compute resources in a namespace and automatically inject them to Containers at runtime.

Enabling LimitRange

LimitRange support has been enabled by default since Kubernetes 1.10.

A LimitRange is enforced in a particular namespace when there is a LimitRange object in that namespace.

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

Overview of Limit Range

• The administrator creates one LimitRange in one namespace.
• Users create resources like Pods, Containers, and PersistentVolumeClaims in the namespace.
• The LimitRanger admission controller enforces defaults and limits for all Pods and Containers that do not set compute resource requirements and tracks usage to ensure it does not exceed resource minimum, maximum and ratio defined in any LimitRange present in the namespace.
• If creating or updating a resource (Pod, Container, PersistentVolumeClaim) that violates a LimitRange constraint, the request to the API server will fail with an HTTP status code 403 FORBIDDEN and a message explaining the constraint that have been violated.
• If a LimitRange is activated in a namespace for compute resources like cpu and memory, users must specify requests or limits for those values. Otherwise, the system may reject Pod creation.
• LimitRange validations occurs only at Pod Admission stage, not on Running Pods.

Examples of policies that could be created using limit range are:

• In a 2 node cluster with a capacity of 8 GiB RAM and 16 cores, constrain Pods in a namespace to request 100m of CPU with a max limit of 500m for CPU and request 200Mi for Memory with a max limit of 600Mi for Memory.
• Define default CPU limit and request to 150m and memory default request to 300Mi for Containers started with no cpu and memory requests in their specs.

In the case where the total limits of the namespace is less than the sum of the limits of the Pods/Containers, there may be contention for resources. In this case, the Containers or Pods will not be created.

Neither contention nor changes to a LimitRange will affect already created resources.

What's next

Refer to the LimitRanger design document for more information.

For examples on using limits, see:

• how to configure minimum and maximum CPU constraints per namespace.
• how to configure minimum and maximum Memory constraints per namespace.
• how to configure default CPU Requests and Limits per namespace.
• how to configure default Memory Requests and Limits per namespace.
• how to configure minimum and maximum Storage consumption per namespace.
• a detailed example on configuring quota per namespace.

2 - Resource Quotas

When several users or teams share a cluster with a fixed number of nodes, there is a concern that one team could use more than its fair share of resources.

Resource quotas are a tool for administrators to address this concern.

A resource quota, defined by a ResourceQuota object, provides constraints that limit aggregate resource consumption per namespace. It can limit the quantity of objects that can be created in a namespace by type, as well as the total amount of compute resources that may be consumed by resources in that namespace.

Resource quotas work like this:

• Different teams work in different namespaces. Currently this is voluntary, but support for making this mandatory via ACLs is planned.

• The administrator creates one ResourceQuota for each namespace.

• Users create resources (pods, services, etc.) in the namespace, and the quota system tracks usage to ensure it does not exceed hard resource limits defined in a ResourceQuota.

• If creating or updating a resource violates a quota constraint, the request will fail with HTTP status code 403 FORBIDDEN with a message explaining the constraint that would have been violated.

• If quota is enabled in a namespace for compute resources like cpu and memory, users must specify requests or limits for those values; otherwise, the quota system may reject pod creation. Hint: Use the LimitRanger admission controller to force defaults for pods that make no compute resource requirements.

  See the walkthrough for an example of how to avoid this problem.

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

Examples of policies that could be created using namespaces and quotas are:

• In a cluster with a capacity of 32 GiB RAM, and 16 cores, let team A use 20 GiB and 10 cores, let B use 10GiB and 4 cores, and hold 2GiB and 2 cores in reserve for future allocation.
• Limit the "testing" namespace to using 1 core and 1GiB RAM. Let the "production" namespace use any amount.

In the case where the total capacity of the cluster is less than the sum of the quotas of the namespaces, there may be contention for resources. This is handled on a first-come-first-served basis.

Neither contention nor changes to quota will affect already created resources.

Enabling Resource Quota

Resource Quota support is enabled by default for many Kubernetes distributions. It is enabled when the API server --enable-admission-plugins= flag has ResourceQuota as one of its arguments.

A resource quota is enforced in a particular namespace when there is a ResourceQuota in that namespace.

Compute Resource Quota

You can limit the total sum of compute resources that can be requested in a given namespace.

The following resource types are supported:

Resource Quota For Extended Resources

In addition to the resources mentioned above, in release 1.10, quota support for extended resources is added.

As overcommit is not allowed for extended resources, it makes no sense to specify both requests and limits for the same extended resource in a quota. So for extended resources, only quota items with prefix requests. is allowed for now.

Take the GPU resource as an example, if the resource name is nvidia.com/gpu, and you want to limit the total number of GPUs requested in a namespace to 4, you can define a quota as follows:

• requests.nvidia.com/gpu: 4

See Viewing and Setting Quotas for more detail information.

Storage Resource Quota

You can limit the total sum of storage resources that can be requested in a given namespace.

In addition, you can limit consumption of storage resources based on associated storage-class.

For example, if an operator wants to quota storage with gold storage class separate from bronze storage class, the operator can define a quota as follows:

• gold.storageclass.storage.k8s.io/requests.storage: 500Gi
• bronze.storageclass.storage.k8s.io/requests.storage: 100Gi

In release 1.8, quota support for local ephemeral storage is added as an alpha feature:

Object Count Quota

You can set quota for the total number of certain resources of all standard, namespaced resource types using the following syntax:

• count/<resource>.<group> for resources from non-core groups
• count/<resource> for resources from the core group

Here is an example set of resources users may want to put under object count quota:

• count/persistentvolumeclaims
• count/services
• count/secrets
• count/configmaps
• count/replicationcontrollers
• count/deployments.apps
• count/replicasets.apps
• count/statefulsets.apps
• count/jobs.batch
• count/cronjobs.batch

The same syntax can be used for custom resources. For example, to create a quota on a widgets custom resource in the example.com API group, use count/widgets.example.com.

When using count/* resource quota, an object is charged against the quota if it exists in server storage. These types of quotas are useful to protect against exhaustion of storage resources. For example, you may want to limit the number of Secrets in a server given their large size. Too many Secrets in a cluster can actually prevent servers and controllers from starting. You can set a quota for Jobs to protect against a poorly configured CronJob. CronJobs that create too many Jobs in a namespace can lead to a denial of service.

It is also possible to do generic object count quota on a limited set of resources. The following types are supported:

For example, pods quota counts and enforces a maximum on the number of pods created in a single namespace that are not terminal. You might want to set a pods quota on a namespace to avoid the case where a user creates many small pods and exhausts the cluster's supply of Pod IPs.

Quota Scopes

Each quota can have an associated set of scopes. A quota will only measure usage for a resource if it matches the intersection of enumerated scopes.

When a scope is added to the quota, it limits the number of resources it supports to those that pertain to the scope. Resources specified on the quota outside of the allowed set results in a validation error.

The BestEffort scope restricts a quota to tracking the following resource:

• pods

The Terminating, NotTerminating, NotBestEffort and PriorityClass scopes restrict a quota to tracking the following resources:

• pods
• cpu
• memory
• requests.cpu
• requests.memory
• limits.cpu
• limits.memory

Note that you cannot specify both the Terminating and the NotTerminating scopes in the same quota, and you cannot specify both the BestEffort and NotBestEffort scopes in the same quota either.

The scopeSelector supports the following values in the operator field:

• In
• NotIn
• Exists
• DoesNotExist

When using one of the following values as the scopeName when defining the scopeSelector, the operator must be Exists.

• Terminating
• NotTerminating
• BestEffort
• NotBestEffort

If the operator is In or NotIn, the values field must have at least one value. For example:

  scopeSelector:
    matchExpressions:
      - scopeName: PriorityClass
        operator: In
        values:
          - middle

If the operator is Exists or DoesNotExist, the values field must NOT be specified.

Resource Quota Per PriorityClass

Pods can be created at a specific priority. You can control a pod's consumption of system resources based on a pod's priority, by using the scopeSelector field in the quota spec.

A quota is matched and consumed only if scopeSelector in the quota spec selects the pod.

When quota is scoped for priority class using scopeSelector field, quota object is restricted to track only following resources:

• pods
• cpu
• memory
• ephemeral-storage
• limits.cpu
• limits.memory
• limits.ephemeral-storage
• requests.cpu
• requests.memory
• requests.ephemeral-storage

This example creates a quota object and matches it with pods at specific priorities. The example works as follows:

• Pods in the cluster have one of the three priority classes, "low", "medium", "high".
• One quota object is created for each priority.

Save the following YAML to a file quota.yml.

apiVersion: v1
kind: List
items:
- apiVersion: v1
  kind: ResourceQuota
  metadata:
    name: pods-high
  spec:
    hard:
      cpu: "1000"
      memory: 200Gi
      pods: "10"
    scopeSelector:
      matchExpressions:
      - operator : In
        scopeName: PriorityClass
        values: ["high"]
- apiVersion: v1
  kind: ResourceQuota
  metadata:
    name: pods-medium
  spec:
    hard:
      cpu: "10"
      memory: 20Gi
      pods: "10"
    scopeSelector:
      matchExpressions:
      - operator : In
        scopeName: PriorityClass
        values: ["medium"]
- apiVersion: v1
  kind: ResourceQuota
  metadata:
    name: pods-low
  spec:
    hard:
      cpu: "5"
      memory: 10Gi
      pods: "10"
    scopeSelector:
      matchExpressions:
      - operator : In
        scopeName: PriorityClass
        values: ["low"]

Apply the YAML using kubectl create.

kubectl create -f ./quota.yml

resourcequota/pods-high created
resourcequota/pods-medium created
resourcequota/pods-low created

Verify that Used quota is 0 using kubectl describe quota.

kubectl describe quota

Name:       pods-high
Namespace:  default
Resource    Used  Hard
--------    ----  ----
cpu         0     1k
memory      0     200Gi
pods        0     10


Name:       pods-low
Namespace:  default
Resource    Used  Hard
--------    ----  ----
cpu         0     5
memory      0     10Gi
pods        0     10


Name:       pods-medium
Namespace:  default
Resource    Used  Hard
--------    ----  ----
cpu         0     10
memory      0     20Gi
pods        0     10

Create a pod with priority "high". Save the following YAML to a file high-priority-pod.yml.

apiVersion: v1
kind: Pod
metadata:
  name: high-priority
spec:
  containers:
  - name: high-priority
    image: ubuntu
    command: ["/bin/sh"]
    args: ["-c", "while true; do echo hello; sleep 10;done"]
    resources:
      requests:
        memory: "10Gi"
        cpu: "500m"
      limits:
        memory: "10Gi"
        cpu: "500m"
  priorityClassName: high

Apply it with kubectl create.

kubectl create -f ./high-priority-pod.yml

Verify that "Used" stats for "high" priority quota, pods-high, has changed and that the other two quotas are unchanged.

kubectl describe quota

Name:       pods-high
Namespace:  default
Resource    Used  Hard
--------    ----  ----
cpu         500m  1k
memory      10Gi  200Gi
pods        1     10


Name:       pods-low
Namespace:  default
Resource    Used  Hard
--------    ----  ----
cpu         0     5
memory      0     10Gi
pods        0     10


Name:       pods-medium
Namespace:  default
Resource    Used  Hard
--------    ----  ----
cpu         0     10
memory      0     20Gi
pods        0     10

Cross-namespace Pod Affinity Quota

Operators can use CrossNamespacePodAffinity quota scope to limit which namespaces are allowed to have pods with affinity terms that cross namespaces. Specifically, it controls which pods are allowed to set namespaces or namespaceSelector fields in pod affinity terms.

Preventing users from using cross-namespace affinity terms might be desired since a pod with anti-affinity constraints can block pods from all other namespaces from getting scheduled in a failure domain.

Using this scope operators can prevent certain namespaces (foo-ns in the example below) from having pods that use cross-namespace pod affinity by creating a resource quota object in that namespace with CrossNamespaceAffinity scope and hard limit of 0:

apiVersion: v1
kind: ResourceQuota
metadata:
  name: disable-cross-namespace-affinity
  namespace: foo-ns
spec:
  hard:
    pods: "0"
  scopeSelector:
    matchExpressions:
    - scopeName: CrossNamespaceAffinity

If operators want to disallow using namespaces and namespaceSelector by default, and only allow it for specific namespaces, they could configure CrossNamespaceAffinity as a limited resource by setting the kube-apiserver flag --admission-control-config-file to the path of the following configuration file:

apiVersion: apiserver.config.k8s.io/v1
kind: AdmissionConfiguration
plugins:
- name: "ResourceQuota"
  configuration:
    apiVersion: apiserver.config.k8s.io/v1
    kind: ResourceQuotaConfiguration
    limitedResources:
    - resource: pods
      matchScopes:
      - scopeName: CrossNamespaceAffinity

With the above configuration, pods can use namespaces and namespaceSelector in pod affinity only if the namespace where they are created have a resource quota object with CrossNamespaceAffinity scope and a hard limit greater than or equal to the number of pods using those fields.

This feature is beta and enabled by default. You can disable it using the feature gate PodAffinityNamespaceSelector in both kube-apiserver and kube-scheduler.

Requests compared to Limits

When allocating compute resources, each container may specify a request and a limit value for either CPU or memory. The quota can be configured to quota either value.

If the quota has a value specified for requests.cpu or requests.memory, then it requires that every incoming container makes an explicit request for those resources. If the quota has a value specified for limits.cpu or limits.memory, then it requires that every incoming container specifies an explicit limit for those resources.

Viewing and Setting Quotas

Kubectl supports creating, updating, and viewing quotas:

kubectl create namespace myspace

cat <<EOF > compute-resources.yaml
apiVersion: v1
kind: ResourceQuota
metadata:
  name: compute-resources
spec:
  hard:
    requests.cpu: "1"
    requests.memory: 1Gi
    limits.cpu: "2"
    limits.memory: 2Gi
    requests.nvidia.com/gpu: 4
EOF

kubectl create -f ./compute-resources.yaml --namespace=myspace

cat <<EOF > object-counts.yaml
apiVersion: v1
kind: ResourceQuota
metadata:
  name: object-counts
spec:
  hard:
    configmaps: "10"
    persistentvolumeclaims: "4"
    pods: "4"
    replicationcontrollers: "20"
    secrets: "10"
    services: "10"
    services.loadbalancers: "2"
EOF

kubectl create -f ./object-counts.yaml --namespace=myspace

kubectl get quota --namespace=myspace

NAME                    AGE
compute-resources       30s
object-counts           32s

kubectl describe quota compute-resources --namespace=myspace

Name:                    compute-resources
Namespace:               myspace
Resource                 Used  Hard
--------                 ----  ----
limits.cpu               0     2
limits.memory            0     2Gi
requests.cpu             0     1
requests.memory          0     1Gi
requests.nvidia.com/gpu  0     4

kubectl describe quota object-counts --namespace=myspace

Name:                   object-counts
Namespace:              myspace
Resource                Used    Hard
--------                ----    ----
configmaps              0       10
persistentvolumeclaims  0       4
pods                    0       4
replicationcontrollers  0       20
secrets                 1       10
services                0       10
services.loadbalancers  0       2

Kubectl also supports object count quota for all standard namespaced resources using the syntax count/<resource>.<group>:

kubectl create namespace myspace

kubectl create quota test --hard=count/deployments.apps=2,count/replicasets.apps=4,count/pods=3,count/secrets=4 --namespace=myspace

kubectl create deployment nginx --image=nginx --namespace=myspace --replicas=2

kubectl describe quota --namespace=myspace

Name:                         test
Namespace:                    myspace
Resource                      Used  Hard
--------                      ----  ----
count/deployments.apps        1     2
count/pods                    2     3
count/replicasets.apps        1     4
count/secrets                 1     4

Quota and Cluster Capacity

ResourceQuotas are independent of the cluster capacity. They are expressed in absolute units. So, if you add nodes to your cluster, this does not automatically give each namespace the ability to consume more resources.

Sometimes more complex policies may be desired, such as:

• Proportionally divide total cluster resources among several teams.
• Allow each tenant to grow resource usage as needed, but have a generous limit to prevent accidental resource exhaustion.
• Detect demand from one namespace, add nodes, and increase quota.

Such policies could be implemented using ResourceQuotas as building blocks, by writing a "controller" that watches the quota usage and adjusts the quota hard limits of each namespace according to other signals.

Note that resource quota divides up aggregate cluster resources, but it creates no restrictions around nodes: pods from several namespaces may run on the same node.

Limit Priority Class consumption by default

It may be desired that pods at a particular priority, eg. "cluster-services", should be allowed in a namespace, if and only if, a matching quota object exists.

With this mechanism, operators are able to restrict usage of certain high priority classes to a limited number of namespaces and not every namespace will be able to consume these priority classes by default.

To enforce this, kube-apiserver flag --admission-control-config-file should be used to pass path to the following configuration file:

apiVersion: apiserver.config.k8s.io/v1
kind: AdmissionConfiguration
plugins:
- name: "ResourceQuota"
  configuration:
    apiVersion: apiserver.config.k8s.io/v1
    kind: ResourceQuotaConfiguration
    limitedResources:
    - resource: pods
      matchScopes:
      - scopeName: PriorityClass
        operator: In
        values: ["cluster-services"]

Then, create a resource quota object in the kube-system namespace:

policy/priority-class-resourcequota.yaml

apiVersion: v1
kind: ResourceQuota
metadata:
  name: pods-cluster-services
spec:
  scopeSelector:
    matchExpressions:
      - operator : In
        scopeName: PriorityClass
        values: ["cluster-services"]

kubectl apply -f https://k8s.io/examples/policy/priority-class-resourcequota.yaml -n kube-system

resourcequota/pods-cluster-services created

In this case, a pod creation will be allowed if:

1. the Pod's priorityClassName is not specified.
2. the Pod's priorityClassName is specified to a value other than cluster-services.
3. the Pod's priorityClassName is set to cluster-services, it is to be created in the kube-system namespace, and it has passed the resource quota check.

A Pod creation request is rejected if its priorityClassName is set to cluster-services and it is to be created in a namespace other than kube-system.

What's next

• See ResourceQuota design doc for more information.
• See a detailed example for how to use resource quota.
• Read Quota support for priority class design doc.
• See LimitedResources

3 - Process ID Limits And Reservations

Kubernetes allow you to limit the number of process IDs (PIDs) that a Pod can use. You can also reserve a number of allocatable PIDs for each node for use by the operating system and daemons (rather than by Pods).

Process IDs (PIDs) are a fundamental resource on nodes. It is trivial to hit the task limit without hitting any other resource limits, which can then cause instability to a host machine.

Cluster administrators require mechanisms to ensure that Pods running in the cluster cannot induce PID exhaustion that prevents host daemons (such as the kubelet or kube-proxy, and potentially also the container runtime) from running. In addition, it is important to ensure that PIDs are limited among Pods in order to ensure they have limited impact on other workloads on the same node.

You can configure a kubelet to limit the number of PIDs a given Pod can consume. For example, if your node's host OS is set to use a maximum of 262144 PIDs and expect to host less than 250 Pods, one can give each Pod a budget of 1000 PIDs to prevent using up that node's overall number of available PIDs. If the admin wants to overcommit PIDs similar to CPU or memory, they may do so as well with some additional risks. Either way, a single Pod will not be able to bring the whole machine down. This kind of resource limiting helps to prevent simple fork bombs from affecting operation of an entire cluster.

Per-Pod PID limiting allows administrators to protect one Pod from another, but does not ensure that all Pods scheduled onto that host are unable to impact the node overall. Per-Pod limiting also does not protect the node agents themselves from PID exhaustion.

You can also reserve an amount of PIDs for node overhead, separate from the allocation to Pods. This is similar to how you can reserve CPU, memory, or other resources for use by the operating system and other facilities outside of Pods and their containers.

PID limiting is a an important sibling to compute resource requests and limits. However, you specify it in a different way: rather than defining a Pod's resource limit in the .spec for a Pod, you configure the limit as a setting on the kubelet. Pod-defined PID limits are not currently supported.

Node PID limits

Kubernetes allows you to reserve a number of process IDs for the system use. To configure the reservation, use the parameter pid=<number> in the --system-reserved and --kube-reserved command line options to the kubelet. The value you specified declares that the specified number of process IDs will be reserved for the system as a whole and for Kubernetes system daemons respectively.

Pod PID limits

Kubernetes allows you to limit the number of processes running in a Pod. You specify this limit at the node level, rather than configuring it as a resource limit for a particular Pod. Each Node can have a different PID limit.
To configure the limit, you can specify the command line parameter --pod-max-pids to the kubelet, or set PodPidsLimit in the kubelet configuration file.

PID based eviction

You can configure kubelet to start terminating a Pod when it is misbehaving and consuming abnormal amount of resources. This feature is called eviction. You can Configure Out of Resource Handling for various eviction signals. Use pid.available eviction signal to configure the threshold for number of PIDs used by Pod. You can set soft and hard eviction policies. However, even with the hard eviction policy, if the number of PIDs growing very fast, node can still get into unstable state by hitting the node PIDs limit. Eviction signal value is calculated periodically and does NOT enforce the limit.

PID limiting - per Pod and per Node sets the hard limit. Once the limit is hit, workload will start experiencing failures when trying to get a new PID. It may or may not lead to rescheduling of a Pod, depending on how workload reacts on these failures and how liveleness and readiness probes are configured for the Pod. However, if limits were set correctly, you can guarantee that other Pods workload and system processes will not run out of PIDs when one Pod is misbehaving.

What's next

• Refer to the PID Limiting enhancement document for more information.
• For historical context, read Process ID Limiting for Stability Improvements in Kubernetes 1.14.
• Read Managing Resources for Containers.
• Learn how to Configure Out of Resource Handling.

4 - Node Resource Managers

In order to support latency-critical and high-throughput workloads, Kubernetes offers a suite of Resource Managers. The managers aim to co-ordinate and optimise node's resources alignment for pods configured with a specific requirement for CPUs, devices, and memory (hugepages) resources.

The main manager, the Topology Manager, is a Kubelet component that co-ordinates the overall resource management process through its policy.

The configuration of individual managers is elaborated in dedicated documents:

• CPU Manager Policies
• Device Manager
• Memory Manager Policies