Scheduling, Preemption and Eviction

• 1: Kubernetes Scheduler
• 2: Assigning Pods to Nodes
• 3: Pod Overhead
• 4: Pod Scheduling Readiness
• 5: Pod Topology Spread Constraints
• 6: Taints and Tolerations
• 7: Scheduling Framework
• 8: Dynamic Resource Allocation
• 9: Scheduler Performance Tuning
• 10: Resource Bin Packing
• 11: Pod Priority and Preemption
• 12: Node-pressure Eviction
• 13: API-initiated Eviction

1 - Kubernetes Scheduler

In Kubernetes, scheduling refers to making sure that Pods are matched to Nodes so that Kubelet can run them.

Scheduling overview

A scheduler watches for newly created Pods that have no Node assigned. For every Pod that the scheduler discovers, the scheduler becomes responsible for finding the best Node for that Pod to run on. The scheduler reaches this placement decision taking into account the scheduling principles described below.

If you want to understand why Pods are placed onto a particular Node, or if you're planning to implement a custom scheduler yourself, this page will help you learn about scheduling.

kube-scheduler

kube-scheduler is the default scheduler for Kubernetes and runs as part of the control plane. kube-scheduler is designed so that, if you want and need to, you can write your own scheduling component and use that instead.

Kube-scheduler selects an optimal node to run newly created or not yet scheduled (unscheduled) pods. Since containers in pods - and pods themselves - can have different requirements, the scheduler filters out any nodes that don't meet a Pod's specific scheduling needs. Alternatively, the API lets you specify a node for a Pod when you create it, but this is unusual and is only done in special cases.

In a cluster, Nodes that meet the scheduling requirements for a Pod are called feasible nodes. If none of the nodes are suitable, the pod remains unscheduled until the scheduler is able to place it.

The scheduler finds feasible Nodes for a Pod and then runs a set of functions to score the feasible Nodes and picks a Node with the highest score among the feasible ones to run the Pod. The scheduler then notifies the API server about this decision in a process called binding.

Factors that need to be taken into account for scheduling decisions include individual and collective resource requirements, hardware / software / policy constraints, affinity and anti-affinity specifications, data locality, inter-workload interference, and so on.

Node selection in kube-scheduler

kube-scheduler selects a node for the pod in a 2-step operation:

1. Filtering
2. Scoring

The filtering step finds the set of Nodes where it's feasible to schedule the Pod. For example, the PodFitsResources filter checks whether a candidate Node has enough available resources to meet a Pod's specific resource requests. After this step, the node list contains any suitable Nodes; often, there will be more than one. If the list is empty, that Pod isn't (yet) schedulable.

In the scoring step, the scheduler ranks the remaining nodes to choose the most suitable Pod placement. The scheduler assigns a score to each Node that survived filtering, basing this score on the active scoring rules.

Finally, kube-scheduler assigns the Pod to the Node with the highest ranking. If there is more than one node with equal scores, kube-scheduler selects one of these at random.

There are two supported ways to configure the filtering and scoring behavior of the scheduler:

1. Scheduling Policies allow you to configure Predicates for filtering and Priorities for scoring.
2. Scheduling Profiles allow you to configure Plugins that implement different scheduling stages, including: QueueSort, Filter, Score, Bind, Reserve, Permit, and others. You can also configure the kube-scheduler to run different profiles.

What's next

• Read about scheduler performance tuning
• Read about Pod topology spread constraints
• Read the reference documentation for kube-scheduler
• Read the kube-scheduler config (v1) reference
• Learn about configuring multiple schedulers
• Learn about topology management policies
• Learn about Pod Overhead
• Learn about scheduling of Pods that use volumes in:
  • Volume Topology Support
  • Storage Capacity Tracking
  • Node-specific Volume Limits

2 - Assigning Pods to Nodes

You can constrain a Pod so that it is restricted to run on particular node(s), or to prefer to run on particular nodes. There are several ways to do this and the recommended approaches all use label selectors to facilitate the selection. Often, you do not need to set any such constraints; the scheduler will automatically do a reasonable placement (for example, spreading your Pods across nodes so as not place Pods on a node with insufficient free resources). However, there are some circumstances where you may want to control which node the Pod deploys to, for example, to ensure that a Pod ends up on a node with an SSD attached to it, or to co-locate Pods from two different services that communicate a lot into the same availability zone.

You can use any of the following methods to choose where Kubernetes schedules specific Pods:

• nodeSelector field matching against node labels
• Affinity and anti-affinity
• nodeName field
• Pod topology spread constraints

Node labels

Like many other Kubernetes objects, nodes have labels. You can attach labels manually. Kubernetes also populates a standard set of labels on all nodes in a cluster.

Node isolation/restriction

Adding labels to nodes allows you to target Pods for scheduling on specific nodes or groups of nodes. You can use this functionality to ensure that specific Pods only run on nodes with certain isolation, security, or regulatory properties.

If you use labels for node isolation, choose label keys that the kubelet cannot modify. This prevents a compromised node from setting those labels on itself so that the scheduler schedules workloads onto the compromised node.

The NodeRestriction admission plugin prevents the kubelet from setting or modifying labels with a node-restriction.kubernetes.io/ prefix.

To make use of that label prefix for node isolation:

1. Ensure you are using the Node authorizer and have enabled the NodeRestriction admission plugin.
2. Add labels with the node-restriction.kubernetes.io/ prefix to your nodes, and use those labels in your node selectors. For example, example.com.node-restriction.kubernetes.io/fips=true or example.com.node-restriction.kubernetes.io/pci-dss=true.

nodeSelector

nodeSelector is the simplest recommended form of node selection constraint. You can add the nodeSelector field to your Pod specification and specify the node labels you want the target node to have. Kubernetes only schedules the Pod onto nodes that have each of the labels you specify.

See Assign Pods to Nodes for more information.

Affinity and anti-affinity

nodeSelector is the simplest way to constrain Pods to nodes with specific labels. Affinity and anti-affinity expands the types of constraints you can define. Some of the benefits of affinity and anti-affinity include:

• The affinity/anti-affinity language is more expressive. nodeSelector only selects nodes with all the specified labels. Affinity/anti-affinity gives you more control over the selection logic.
• You can indicate that a rule is soft or preferred, so that the scheduler still schedules the Pod even if it can't find a matching node.
• You can constrain a Pod using labels on other Pods running on the node (or other topological domain), instead of just node labels, which allows you to define rules for which Pods can be co-located on a node.

The affinity feature consists of two types of affinity:

• Node affinity functions like the nodeSelector field but is more expressive and allows you to specify soft rules.
• Inter-pod affinity/anti-affinity allows you to constrain Pods against labels on other Pods.

Node affinity

Node affinity is conceptually similar to nodeSelector, allowing you to constrain which nodes your Pod can be scheduled on based on node labels. There are two types of node affinity:

• requiredDuringSchedulingIgnoredDuringExecution: The scheduler can't schedule the Pod unless the rule is met. This functions like nodeSelector, but with a more expressive syntax.
• preferredDuringSchedulingIgnoredDuringExecution: The scheduler tries to find a node that meets the rule. If a matching node is not available, the scheduler still schedules the Pod.

You can specify node affinities using the .spec.affinity.nodeAffinity field in your Pod spec.

For example, consider the following Pod spec:

pods/pod-with-node-affinity.yaml

apiVersion: v1
kind: Pod
metadata:
  name: with-node-affinity
spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: topology.kubernetes.io/zone
            operator: In
            values:
            - antarctica-east1
            - antarctica-west1
      preferredDuringSchedulingIgnoredDuringExecution:
      - weight: 1
        preference:
          matchExpressions:
          - key: another-node-label-key
            operator: In
            values:
            - another-node-label-value
  containers:
  - name: with-node-affinity
    image: registry.k8s.io/pause:2.0

In this example, the following rules apply:

• The node must have a label with the key topology.kubernetes.io/zone and the value of that label must be either antarctica-east1 or antarctica-west1.
• The node preferably has a label with the key another-node-label-key and the value another-node-label-value.

You can use the operator field to specify a logical operator for Kubernetes to use when interpreting the rules. You can use In, NotIn, Exists, DoesNotExist, Gt and Lt.

Read Operators to learn more about how these work.

NotIn and DoesNotExist allow you to define node anti-affinity behavior. Alternatively, you can use node taints to repel Pods from specific nodes.

See Assign Pods to Nodes using Node Affinity for more information.

Node affinity weight

You can specify a weight between 1 and 100 for each instance of the preferredDuringSchedulingIgnoredDuringExecution affinity type. When the scheduler finds nodes that meet all the other scheduling requirements of the Pod, the scheduler iterates through every preferred rule that the node satisfies and adds the value of the weight for that expression to a sum.

The final sum is added to the score of other priority functions for the node. Nodes with the highest total score are prioritized when the scheduler makes a scheduling decision for the Pod.

For example, consider the following Pod spec:

pods/pod-with-affinity-preferred-weight.yaml

apiVersion: v1
kind: Pod
metadata:
  name: with-affinity-preferred-weight
spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: kubernetes.io/os
            operator: In
            values:
            - linux
      preferredDuringSchedulingIgnoredDuringExecution:
      - weight: 1
        preference:
          matchExpressions:
          - key: label-1
            operator: In
            values:
            - key-1
      - weight: 50
        preference:
          matchExpressions:
          - key: label-2
            operator: In
            values:
            - key-2
  containers:
  - name: with-node-affinity
    image: registry.k8s.io/pause:2.0

If there are two possible nodes that match the preferredDuringSchedulingIgnoredDuringExecution rule, one with the label-1:key-1 label and another with the label-2:key-2 label, the scheduler considers the weight of each node and adds the weight to the other scores for that node, and schedules the Pod onto the node with the highest final score.

Node affinity per scheduling profile

When configuring multiple scheduling profiles, you can associate a profile with a node affinity, which is useful if a profile only applies to a specific set of nodes. To do so, add an addedAffinity to the args field of the NodeAffinity plugin in the scheduler configuration. For example:

apiVersion: kubescheduler.config.k8s.io/v1beta3
kind: KubeSchedulerConfiguration

profiles:
  - schedulerName: default-scheduler
  - schedulerName: foo-scheduler
    pluginConfig:
      - name: NodeAffinity
        args:
          addedAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
              nodeSelectorTerms:
              - matchExpressions:
                - key: scheduler-profile
                  operator: In
                  values:
                  - foo

The addedAffinity is applied to all Pods that set .spec.schedulerName to foo-scheduler, in addition to the NodeAffinity specified in the PodSpec. That is, in order to match the Pod, nodes need to satisfy addedAffinity and the Pod's .spec.NodeAffinity.

Since the addedAffinity is not visible to end users, its behavior might be unexpected to them. Use node labels that have a clear correlation to the scheduler profile name.

Inter-pod affinity and anti-affinity

Inter-pod affinity and anti-affinity allow you to constrain which nodes your Pods can be scheduled on based on the labels of Pods already running on that node, instead of the node labels.

Inter-pod affinity and anti-affinity rules take the form "this Pod should (or, in the case of anti-affinity, should not) run in an X if that X is already running one or more Pods that meet rule Y", where X is a topology domain like node, rack, cloud provider zone or region, or similar and Y is the rule Kubernetes tries to satisfy.

You express these rules (Y) as label selectors with an optional associated list of namespaces. Pods are namespaced objects in Kubernetes, so Pod labels also implicitly have namespaces. Any label selectors for Pod labels should specify the namespaces in which Kubernetes should look for those labels.

You express the topology domain (X) using a topologyKey, which is the key for the node label that the system uses to denote the domain. For examples, see Well-Known Labels, Annotations and Taints.

Types of inter-pod affinity and anti-affinity

Similar to node affinity are two types of Pod affinity and anti-affinity as follows:

• requiredDuringSchedulingIgnoredDuringExecution
• preferredDuringSchedulingIgnoredDuringExecution

For example, you could use requiredDuringSchedulingIgnoredDuringExecution affinity to tell the scheduler to co-locate Pods of two services in the same cloud provider zone because they communicate with each other a lot. Similarly, you could use preferredDuringSchedulingIgnoredDuringExecution anti-affinity to spread Pods from a service across multiple cloud provider zones.

To use inter-pod affinity, use the affinity.podAffinity field in the Pod spec. For inter-pod anti-affinity, use the affinity.podAntiAffinity field in the Pod spec.

Scheduling a group of pods with inter-pod affinity to themselves

If the current Pod being scheduled is the first in a series that have affinity to themselves, it is allowed to be scheduled if it passes all other affinity checks. This is determined by verifying that no other pod in the cluster matches the namespace and selector of this pod, that the pod matches its own terms, and the chosen node matches all requested topologies. This ensures that there will not be a deadlock even if all the pods have inter-pod affinity specified.

Pod affinity example

Consider the following Pod spec:

pods/pod-with-pod-affinity.yaml

apiVersion: v1
kind: Pod
metadata:
  name: with-pod-affinity
spec:
  affinity:
    podAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
      - labelSelector:
          matchExpressions:
          - key: security
            operator: In
            values:
            - S1
        topologyKey: topology.kubernetes.io/zone
    podAntiAffinity:
      preferredDuringSchedulingIgnoredDuringExecution:
      - weight: 100
        podAffinityTerm:
          labelSelector:
            matchExpressions:
            - key: security
              operator: In
              values:
              - S2
          topologyKey: topology.kubernetes.io/zone
  containers:
  - name: with-pod-affinity
    image: registry.k8s.io/pause:2.0

This example defines one Pod affinity rule and one Pod anti-affinity rule. The Pod affinity rule uses the "hard" requiredDuringSchedulingIgnoredDuringExecution, while the anti-affinity rule uses the "soft" preferredDuringSchedulingIgnoredDuringExecution.

The affinity rule specifies that the scheduler is allowed to place the example Pod on a node only if that node belongs to a specific zone where other Pods have been labeled with security=S1. For instance, if we have a cluster with a designated zone, let's call it "Zone V," consisting of nodes labeled with topology.kubernetes.io/zone=V, the scheduler can assign the Pod to any node within Zone V, as long as there is at least one Pod within Zone V already labeled with security=S1. Conversely, if there are no Pods with security=S1 labels in Zone V, the scheduler will not assign the example Pod to any node in that zone.

The anti-affinity rule specifies that the scheduler should try to avoid scheduling the Pod on a node if that node belongs to a specific zone where other Pods have been labeled with security=S2. For instance, if we have a cluster with a designated zone, let's call it "Zone R," consisting of nodes labeled with topology.kubernetes.io/zone=R, the scheduler should avoid assigning the Pod to any node within Zone R, as long as there is at least one Pod within Zone R already labeled with security=S2. Conversely, the anti-affinity rule does not impact scheduling into Zone R if there are no Pods with security=S2 labels.

To get yourself more familiar with the examples of Pod affinity and anti-affinity, refer to the design proposal.

You can use the In, NotIn, Exists and DoesNotExist values in the operator field for Pod affinity and anti-affinity.

Read Operators to learn more about how these work.

In principle, the topologyKey can be any allowed label key with the following exceptions for performance and security reasons:

• For Pod affinity and anti-affinity, an empty topologyKey field is not allowed in both requiredDuringSchedulingIgnoredDuringExecution and preferredDuringSchedulingIgnoredDuringExecution.
• For requiredDuringSchedulingIgnoredDuringExecution Pod anti-affinity rules, the admission controller LimitPodHardAntiAffinityTopology limits topologyKey to kubernetes.io/hostname. You can modify or disable the admission controller if you want to allow custom topologies.

In addition to labelSelector and topologyKey, you can optionally specify a list of namespaces which the labelSelector should match against using the namespaces field at the same level as labelSelector and topologyKey. If omitted or empty, namespaces defaults to the namespace of the Pod where the affinity/anti-affinity definition appears.

Namespace selector

You can also select matching namespaces using namespaceSelector, which is a label query over the set of namespaces. The affinity term is applied to namespaces selected by both namespaceSelector and the namespaces field. Note that an empty namespaceSelector ({}) matches all namespaces, while a null or empty namespaces list and null namespaceSelector matches the namespace of the Pod where the rule is defined.

matchLabelKeys

Kubernetes includes an optional matchLabelKeys field for Pod affinity or anti-affinity. The field specifies keys for the labels that should match with the incoming Pod's labels, when satisfying the Pod (anti)affinity.

The keys are used to look up values from the pod labels; those key-value labels are combined (using AND) with the match restrictions defined using the labelSelector field. The combined filtering selects the set of existing pods that will be taken into Pod (anti)affinity calculation.

A common use case is to use matchLabelKeys with pod-template-hash (set on Pods managed as part of a Deployment, where the value is unique for each revision). Using pod-template-hash in matchLabelKeys allows you to target the Pods that belong to the same revision as the incoming Pod, so that a rolling upgrade won't break affinity.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: application-server
...
spec:
  template:
    spec:
      affinity:
        podAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
          - labelSelector:
              matchExpressions:
              - key: app
                operator: In
                values:
                - database
            topologyKey: topology.kubernetes.io/zone
            # Only Pods from a given rollout are taken into consideration when calculating pod affinity.
            # If you update the Deployment, the replacement Pods follow their own affinity rules
            # (if there are any defined in the new Pod template)
            matchLabelKeys:
            - pod-template-hash

mismatchLabelKeys

Kubernetes includes an optional mismatchLabelKeys field for Pod affinity or anti-affinity. The field specifies keys for the labels that should not match with the incoming Pod's labels, when satisfying the Pod (anti)affinity.

One example use case is to ensure Pods go to the topology domain (node, zone, etc) where only Pods from the same tenant or team are scheduled in. In other words, you want to avoid running Pods from two different tenants on the same topology domain at the same time.

apiVersion: v1
kind: Pod
metadata:
  labels:
    # Assume that all relevant Pods have a "tenant" label set
    tenant: tenant-a
...
spec:
  affinity:
    podAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
      # ensure that pods associated with this tenant land on the correct node pool
      - matchLabelKeys:
          - tenant
        topologyKey: node-pool
    podAntiAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
      # ensure that pods associated with this tenant can't schedule to nodes used for another tenant
      - mismatchLabelKeys:
        - tenant # whatever the value of the "tenant" label for this Pod, prevent
                 # scheduling to nodes in any pool where any Pod from a different
                 # tenant is running.
        labelSelector:
          # We have to have the labelSelector which selects only Pods with the tenant label,
          # otherwise this Pod would hate Pods from daemonsets as well, for example,
          # which aren't supposed to have the tenant label.
          matchExpressions:
          - key: tenant
            operator: Exists
        topologyKey: node-pool

More practical use-cases

Inter-pod affinity and anti-affinity can be even more useful when they are used with higher level collections such as ReplicaSets, StatefulSets, Deployments, etc. These rules allow you to configure that a set of workloads should be co-located in the same defined topology; for example, preferring to place two related Pods onto the same node.

For example: imagine a three-node cluster. You use the cluster to run a web application and also an in-memory cache (such as Redis). For this example, also assume that latency between the web application and the memory cache should be as low as is practical. You could use inter-pod affinity and anti-affinity to co-locate the web servers with the cache as much as possible.

In the following example Deployment for the Redis cache, the replicas get the label app=store. The podAntiAffinity rule tells the scheduler to avoid placing multiple replicas with the app=store label on a single node. This creates each cache in a separate node.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: redis-cache
spec:
  selector:
    matchLabels:
      app: store
  replicas: 3
  template:
    metadata:
      labels:
        app: store
    spec:
      affinity:
        podAntiAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
          - labelSelector:
              matchExpressions:
              - key: app
                operator: In
                values:
                - store
            topologyKey: "kubernetes.io/hostname"
      containers:
      - name: redis-server
        image: redis:3.2-alpine

The following example Deployment for the web servers creates replicas with the label app=web-store. The Pod affinity rule tells the scheduler to place each replica on a node that has a Pod with the label app=store. The Pod anti-affinity rule tells the scheduler never to place multiple app=web-store servers on a single node.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: web-server
spec:
  selector:
    matchLabels:
      app: web-store
  replicas: 3
  template:
    metadata:
      labels:
        app: web-store
    spec:
      affinity:
        podAntiAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
          - labelSelector:
              matchExpressions:
              - key: app
                operator: In
                values:
                - web-store
            topologyKey: "kubernetes.io/hostname"
        podAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
          - labelSelector:
              matchExpressions:
              - key: app
                operator: In
                values:
                - store
            topologyKey: "kubernetes.io/hostname"
      containers:
      - name: web-app
        image: nginx:1.16-alpine

Creating the two preceding Deployments results in the following cluster layout, where each web server is co-located with a cache, on three separate nodes.

The overall effect is that each cache instance is likely to be accessed by a single client that is running on the same node. This approach aims to minimize both skew (imbalanced load) and latency.

You might have other reasons to use Pod anti-affinity. See the ZooKeeper tutorial for an example of a StatefulSet configured with anti-affinity for high availability, using the same technique as this example.

nodeName

nodeName is a more direct form of node selection than affinity or nodeSelector. nodeName is a field in the Pod spec. If the nodeName field is not empty, the scheduler ignores the Pod and the kubelet on the named node tries to place the Pod on that node. Using nodeName overrules using nodeSelector or affinity and anti-affinity rules.

Some of the limitations of using nodeName to select nodes are:

• If the named node does not exist, the Pod will not run, and in some cases may be automatically deleted.
• If the named node does not have the resources to accommodate the Pod, the Pod will fail and its reason will indicate why, for example OutOfmemory or OutOfcpu.
• Node names in cloud environments are not always predictable or stable.

Here is an example of a Pod spec using the nodeName field:

apiVersion: v1
kind: Pod
metadata:
  name: nginx
spec:
  containers:
  - name: nginx
    image: nginx
  nodeName: kube-01

The above Pod will only run on the node kube-01.

Pod topology spread constraints

You can use topology spread constraints to control how Pods are spread across your cluster among failure-domains such as regions, zones, nodes, or among any other topology domains that you define. You might do this to improve performance, expected availability, or overall utilization.

Read Pod topology spread constraints to learn more about how these work.

Operators

The following are all the logical operators that you can use in the operator field for nodeAffinity and podAffinity mentioned above.

The following operators can only be used with nodeAffinity.

What's next

• Read more about taints and tolerations.
• Read the design docs for node affinity and for inter-pod affinity/anti-affinity.
• Learn about how the topology manager takes part in node-level resource allocation decisions.
• Learn how to use nodeSelector.
• Learn how to use affinity and anti-affinity.

3 - Pod Overhead

When you run a Pod on a Node, the Pod itself takes an amount of system resources. These resources are additional to the resources needed to run the container(s) inside the Pod. In Kubernetes, Pod Overhead is a way to account for the resources consumed by the Pod infrastructure on top of the container requests & limits.

In Kubernetes, the Pod's overhead is set at admission time according to the overhead associated with the Pod's RuntimeClass.

A pod's overhead is considered in addition to the sum of container resource requests when scheduling a Pod. Similarly, the kubelet will include the Pod overhead when sizing the Pod cgroup, and when carrying out Pod eviction ranking.

Configuring Pod overhead

You need to make sure a RuntimeClass is utilized which defines the overhead field.

Usage example

To work with Pod overhead, you need a RuntimeClass that defines the overhead field. As an example, you could use the following RuntimeClass definition with a virtualization container runtime (in this example, Kata Containers combined with the Firecracker virtual machine monitor) that uses around 120MiB per Pod for the virtual machine and the guest OS:

# You need to change this example to match the actual runtime name, and per-Pod
# resource overhead, that the container runtime is adding in your cluster.
apiVersion: node.k8s.io/v1
kind: RuntimeClass
metadata:
  name: kata-fc
handler: kata-fc
overhead:
  podFixed:
    memory: "120Mi"
    cpu: "250m"

Workloads which are created which specify the kata-fc RuntimeClass handler will take the memory and cpu overheads into account for resource quota calculations, node scheduling, as well as Pod cgroup sizing.

Consider running the given example workload, test-pod:

apiVersion: v1
kind: Pod
metadata:
  name: test-pod
spec:
  runtimeClassName: kata-fc
  containers:
  - name: busybox-ctr
    image: busybox:1.28
    stdin: true
    tty: true
    resources:
      limits:
        cpu: 500m
        memory: 100Mi
  - name: nginx-ctr
    image: nginx
    resources:
      limits:
        cpu: 1500m
        memory: 100Mi

At admission time the RuntimeClass admission controller updates the workload's PodSpec to include the overhead as described in the RuntimeClass. If the PodSpec already has this field defined, the Pod will be rejected. In the given example, since only the RuntimeClass name is specified, the admission controller mutates the Pod to include an overhead.

After the RuntimeClass admission controller has made modifications, you can check the updated Pod overhead value:

kubectl get pod test-pod -o jsonpath='{.spec.overhead}'

The output is:

map[cpu:250m memory:120Mi]

If a ResourceQuota is defined, the sum of container requests as well as the overhead field are counted.

When the kube-scheduler is deciding which node should run a new Pod, the scheduler considers that Pod's overhead as well as the sum of container requests for that Pod. For this example, the scheduler adds the requests and the overhead, then looks for a node that has 2.25 CPU and 320 MiB of memory available.

Once a Pod is scheduled to a node, the kubelet on that node creates a new cgroup for the Pod. It is within this pod that the underlying container runtime will create containers.

If the resource has a limit defined for each container (Guaranteed QoS or Burstable QoS with limits defined), the kubelet will set an upper limit for the pod cgroup associated with that resource (cpu.cfs_quota_us for CPU and memory.limit_in_bytes memory). This upper limit is based on the sum of the container limits plus the overhead defined in the PodSpec.

For CPU, if the Pod is Guaranteed or Burstable QoS, the kubelet will set cpu.shares based on the sum of container requests plus the overhead defined in the PodSpec.

Looking at our example, verify the container requests for the workload:

kubectl get pod test-pod -o jsonpath='{.spec.containers[*].resources.limits}'

The total container requests are 2000m CPU and 200MiB of memory:

map[cpu: 500m memory:100Mi] map[cpu:1500m memory:100Mi]

Check this against what is observed by the node:

kubectl describe node | grep test-pod -B2

The output shows requests for 2250m CPU, and for 320MiB of memory. The requests include Pod overhead:

  Namespace    Name       CPU Requests  CPU Limits   Memory Requests  Memory Limits  AGE
  ---------    ----       ------------  ----------   ---------------  -------------  ---
  default      test-pod   2250m (56%)   2250m (56%)  320Mi (1%)       320Mi (1%)     36m

Verify Pod cgroup limits

Check the Pod's memory cgroups on the node where the workload is running. In the following example, crictl is used on the node, which provides a CLI for CRI-compatible container runtimes. This is an advanced example to show Pod overhead behavior, and it is not expected that users should need to check cgroups directly on the node.

First, on the particular node, determine the Pod identifier:

# Run this on the node where the Pod is scheduled
POD_ID="$(sudo crictl pods --name test-pod -q)"

From this, you can determine the cgroup path for the Pod:

# Run this on the node where the Pod is scheduled
sudo crictl inspectp -o=json $POD_ID | grep cgroupsPath

The resulting cgroup path includes the Pod's pause container. The Pod level cgroup is one directory above.

  "cgroupsPath": "/kubepods/podd7f4b509-cf94-4951-9417-d1087c92a5b2/7ccf55aee35dd16aca4189c952d83487297f3cd760f1bbf09620e206e7d0c27a"

In this specific case, the pod cgroup path is kubepods/podd7f4b509-cf94-4951-9417-d1087c92a5b2. Verify the Pod level cgroup setting for memory:

# Run this on the node where the Pod is scheduled.
# Also, change the name of the cgroup to match the cgroup allocated for your pod.
 cat /sys/fs/cgroup/memory/kubepods/podd7f4b509-cf94-4951-9417-d1087c92a5b2/memory.limit_in_bytes

This is 320 MiB, as expected:

335544320

Observability

Some kube_pod_overhead_* metrics are available in kube-state-metrics to help identify when Pod overhead is being utilized and to help observe stability of workloads running with a defined overhead.

What's next

• Learn more about RuntimeClass
• Read the PodOverhead Design enhancement proposal for extra context

4 - Pod Scheduling Readiness

Pods were considered ready for scheduling once created. Kubernetes scheduler does its due diligence to find nodes to place all pending Pods. However, in a real-world case, some Pods may stay in a "miss-essential-resources" state for a long period. These Pods actually churn the scheduler (and downstream integrators like Cluster AutoScaler) in an unnecessary manner.

By specifying/removing a Pod's .spec.schedulingGates, you can control when a Pod is ready to be considered for scheduling.

Configuring Pod schedulingGates

The schedulingGates field contains a list of strings, and each string literal is perceived as a criteria that Pod should be satisfied before considered schedulable. This field can be initialized only when a Pod is created (either by the client, or mutated during admission). After creation, each schedulingGate can be removed in arbitrary order, but addition of a new scheduling gate is disallowed.

Usage example

To mark a Pod not-ready for scheduling, you can create it with one or more scheduling gates like this:

pods/pod-with-scheduling-gates.yaml

apiVersion: v1
kind: Pod
metadata:
  name: test-pod
spec:
  schedulingGates:
  - name: example.com/foo
  - name: example.com/bar
  containers:
  - name: pause
    image: registry.k8s.io/pause:3.6

After the Pod's creation, you can check its state using:

kubectl get pod test-pod

The output reveals it's in SchedulingGated state:

NAME       READY   STATUS            RESTARTS   AGE
test-pod   0/1     SchedulingGated   0          7s

You can also check its schedulingGates field by running:

kubectl get pod test-pod -o jsonpath='{.spec.schedulingGates}'

The output is:

[{"name":"example.com/foo"},{"name":"example.com/bar"}]

To inform scheduler this Pod is ready for scheduling, you can remove its schedulingGates entirely by reapplying a modified manifest:

pods/pod-without-scheduling-gates.yaml

apiVersion: v1
kind: Pod
metadata:
  name: test-pod
spec:
  containers:
  - name: pause
    image: registry.k8s.io/pause:3.6

You can check if the schedulingGates is cleared by running:

kubectl get pod test-pod -o jsonpath='{.spec.schedulingGates}'

The output is expected to be empty. And you can check its latest status by running:

kubectl get pod test-pod -o wide

Given the test-pod doesn't request any CPU/memory resources, it's expected that this Pod's state get transited from previous SchedulingGated to Running:

NAME       READY   STATUS    RESTARTS   AGE   IP         NODE
test-pod   1/1     Running   0          15s   10.0.0.4   node-2

Observability

The metric scheduler_pending_pods comes with a new label "gated" to distinguish whether a Pod has been tried scheduling but claimed as unschedulable, or explicitly marked as not ready for scheduling. You can use scheduler_pending_pods{queue="gated"} to check the metric result.

Mutable Pod scheduling directives

You can mutate scheduling directives of Pods while they have scheduling gates, with certain constraints. At a high level, you can only tighten the scheduling directives of a Pod. In other words, the updated directives would cause the Pods to only be able to be scheduled on a subset of the nodes that it would previously match. More concretely, the rules for updating a Pod's scheduling directives are as follows:

1. For .spec.nodeSelector, only additions are allowed. If absent, it will be allowed to be set.

2. For spec.affinity.nodeAffinity, if nil, then setting anything is allowed.

3. If NodeSelectorTerms was empty, it will be allowed to be set. If not empty, then only additions of NodeSelectorRequirements to matchExpressions or fieldExpressions are allowed, and no changes to existing matchExpressions and fieldExpressions will be allowed. This is because the terms in .requiredDuringSchedulingIgnoredDuringExecution.NodeSelectorTerms, are ORed while the expressions in nodeSelectorTerms[].matchExpressions and nodeSelectorTerms[].fieldExpressions are ANDed.

4. For .preferredDuringSchedulingIgnoredDuringExecution, all updates are allowed. This is because preferred terms are not authoritative, and so policy controllers don't validate those terms.

What's next

• Read the PodSchedulingReadiness KEP for more details

5 - Pod Topology Spread Constraints

You can use topology spread constraints to control how Pods are spread across your cluster among failure-domains such as regions, zones, nodes, and other user-defined topology domains. This can help to achieve high availability as well as efficient resource utilization.

You can set cluster-level constraints as a default, or configure topology spread constraints for individual workloads.

Motivation

Imagine that you have a cluster of up to twenty nodes, and you want to run a workload that automatically scales how many replicas it uses. There could be as few as two Pods or as many as fifteen. When there are only two Pods, you'd prefer not to have both of those Pods run on the same node: you would run the risk that a single node failure takes your workload offline.

In addition to this basic usage, there are some advanced usage examples that enable your workloads to benefit on high availability and cluster utilization.

As you scale up and run more Pods, a different concern becomes important. Imagine that you have three nodes running five Pods each. The nodes have enough capacity to run that many replicas; however, the clients that interact with this workload are split across three different datacenters (or infrastructure zones). Now you have less concern about a single node failure, but you notice that latency is higher than you'd like, and you are paying for network costs associated with sending network traffic between the different zones.

You decide that under normal operation you'd prefer to have a similar number of replicas scheduled into each infrastructure zone, and you'd like the cluster to self-heal in the case that there is a problem.

Pod topology spread constraints offer you a declarative way to configure that.

topologySpreadConstraints field

The Pod API includes a field, spec.topologySpreadConstraints. The usage of this field looks like the following:

---
apiVersion: v1
kind: Pod
metadata:
  name: example-pod
spec:
  # Configure a topology spread constraint
  topologySpreadConstraints:
    - maxSkew: <integer>
      minDomains: <integer> # optional
      topologyKey: <string>
      whenUnsatisfiable: <string>
      labelSelector: <object>
      matchLabelKeys: <list> # optional; beta since v1.27
      nodeAffinityPolicy: [Honor|Ignore] # optional; beta since v1.26
      nodeTaintsPolicy: [Honor|Ignore] # optional; beta since v1.26
  ### other Pod fields go here

You can read more about this field by running kubectl explain Pod.spec.topologySpreadConstraints or refer to the scheduling section of the API reference for Pod.

Spread constraint definition

You can define one or multiple topologySpreadConstraints entries to instruct the kube-scheduler how to place each incoming Pod in relation to the existing Pods across your cluster. Those fields are:

• maxSkew describes the degree to which Pods may be unevenly distributed. You must specify this field and the number must be greater than zero. Its semantics differ according to the value of whenUnsatisfiable:

  • if you select whenUnsatisfiable: DoNotSchedule, then maxSkew defines the maximum permitted difference between the number of matching pods in the target topology and the global minimum (the minimum number of matching pods in an eligible domain or zero if the number of eligible domains is less than MinDomains). For example, if you have 3 zones with 2, 2 and 1 matching pods respectively, MaxSkew is set to 1 then the global minimum is 1.
  • if you select whenUnsatisfiable: ScheduleAnyway, the scheduler gives higher precedence to topologies that would help reduce the skew.

• minDomains indicates a minimum number of eligible domains. This field is optional. A domain is a particular instance of a topology. An eligible domain is a domain whose nodes match the node selector.

  Note:

  Before Kubernetes v1.30, the minDomains field was only available if the MinDomainsInPodTopologySpread feature gate was enabled (default since v1.28). In older Kubernetes clusters it might be explicitly disabled or the field might not be available.
  • The value of minDomains must be greater than 0, when specified. You can only specify minDomains in conjunction with whenUnsatisfiable: DoNotSchedule.
  • When the number of eligible domains with match topology keys is less than minDomains, Pod topology spread treats global minimum as 0, and then the calculation of skew is performed. The global minimum is the minimum number of matching Pods in an eligible domain, or zero if the number of eligible domains is less than minDomains.
  • When the number of eligible domains with matching topology keys equals or is greater than minDomains, this value has no effect on scheduling.
  • If you do not specify minDomains, the constraint behaves as if minDomains is 1.

• topologyKey is the key of node labels. Nodes that have a label with this key and identical values are considered to be in the same topology. We call each instance of a topology (in other words, a <key, value> pair) a domain. The scheduler will try to put a balanced number of pods into each domain. Also, we define an eligible domain as a domain whose nodes meet the requirements of nodeAffinityPolicy and nodeTaintsPolicy.

• whenUnsatisfiable indicates how to deal with a Pod if it doesn't satisfy the spread constraint:

  • DoNotSchedule (default) tells the scheduler not to schedule it.
  • ScheduleAnyway tells the scheduler to still schedule it while prioritizing nodes that minimize the skew.

• labelSelector is used to find matching Pods. Pods that match this label selector are counted to determine the number of Pods in their corresponding topology domain. See Label Selectors for more details.

• matchLabelKeys is a list of pod label keys to select the pods over which spreading will be calculated. The keys are used to lookup values from the pod labels, those key-value labels are ANDed with labelSelector to select the group of existing pods over which spreading will be calculated for the incoming pod. The same key is forbidden to exist in both matchLabelKeys and labelSelector. matchLabelKeys cannot be set when labelSelector isn't set. Keys that don't exist in the pod labels will be ignored. A null or empty list means only match against the labelSelector.

  With matchLabelKeys, you don't need to update the pod.spec between different revisions. The controller/operator just needs to set different values to the same label key for different revisions. The scheduler will assume the values automatically based on matchLabelKeys. For example, if you are configuring a Deployment, you can use the label keyed with pod-template-hash, which is added automatically by the Deployment controller, to distinguish between different revisions in a single Deployment.

      topologySpreadConstraints:
          - maxSkew: 1
            topologyKey: kubernetes.io/hostname
            whenUnsatisfiable: DoNotSchedule
            labelSelector:
              matchLabels:
                app: foo
            matchLabelKeys:
              - pod-template-hash
  

  Note:

  The matchLabelKeys field is a beta-level field and enabled by default in 1.27. You can disable it by disabling the MatchLabelKeysInPodTopologySpread feature gate.

• nodeAffinityPolicy indicates how we will treat Pod's nodeAffinity/nodeSelector when calculating pod topology spread skew. Options are:

  • Honor: only nodes matching nodeAffinity/nodeSelector are included in the calculations.
  • Ignore: nodeAffinity/nodeSelector are ignored. All nodes are included in the calculations.

  If this value is null, the behavior is equivalent to the Honor policy.

  Note:

  The nodeAffinityPolicy is a beta-level field and enabled by default in 1.26. You can disable it by disabling the NodeInclusionPolicyInPodTopologySpread feature gate.

• nodeTaintsPolicy indicates how we will treat node taints when calculating pod topology spread skew. Options are:

  • Honor: nodes without taints, along with tainted nodes for which the incoming pod has a toleration, are included.
  • Ignore: node taints are ignored. All nodes are included.

  If this value is null, the behavior is equivalent to the Ignore policy.

  Note:

  The nodeTaintsPolicy is a beta-level field and enabled by default in 1.26. You can disable it by disabling the NodeInclusionPolicyInPodTopologySpread feature gate.

When a Pod defines more than one topologySpreadConstraint, those constraints are combined using a logical AND operation: the kube-scheduler looks for a node for the incoming Pod that satisfies all the configured constraints.

Node labels

Topology spread constraints rely on node labels to identify the topology domain(s) that each node is in. For example, a node might have labels:

  region: us-east-1
  zone: us-east-1a

Suppose you have a 4-node cluster with the following labels:

NAME    STATUS   ROLES    AGE     VERSION   LABELS
node1   Ready    <none>   4m26s   v1.16.0   node=node1,zone=zoneA
node2   Ready    <none>   3m58s   v1.16.0   node=node2,zone=zoneA
node3   Ready    <none>   3m17s   v1.16.0   node=node3,zone=zoneB
node4   Ready    <none>   2m43s   v1.16.0   node=node4,zone=zoneB

Then the cluster is logically viewed as below:

Consistency

You should set the same Pod topology spread constraints on all pods in a group.

Usually, if you are using a workload controller such as a Deployment, the pod template takes care of this for you. If you mix different spread constraints then Kubernetes follows the API definition of the field; however, the behavior is more likely to become confusing and troubleshooting is less straightforward.

You need a mechanism to ensure that all the nodes in a topology domain (such as a cloud provider region) are labeled consistently. To avoid you needing to manually label nodes, most clusters automatically populate well-known labels such as kubernetes.io/hostname. Check whether your cluster supports this.

Topology spread constraint examples

Example: one topology spread constraint

Suppose you have a 4-node cluster where 3 Pods labeled foo: bar are located in node1, node2 and node3 respectively:

If you want an incoming Pod to be evenly spread with existing Pods across zones, you can use a manifest similar to:

pods/topology-spread-constraints/one-constraint.yaml

kind: Pod
apiVersion: v1
metadata:
  name: mypod
  labels:
    foo: bar
spec:
  topologySpreadConstraints:
  - maxSkew: 1
    topologyKey: zone
    whenUnsatisfiable: DoNotSchedule
    labelSelector:
      matchLabels:
        foo: bar
  containers:
  - name: pause
    image: registry.k8s.io/pause:3.1

From that manifest, topologyKey: zone implies the even distribution will only be applied to nodes that are labeled zone: <any value> (nodes that don't have a zone label are skipped). The field whenUnsatisfiable: DoNotSchedule tells the scheduler to let the incoming Pod stay pending if the scheduler can't find a way to satisfy the constraint.

If the scheduler placed this incoming Pod into zone A, the distribution of Pods would become [3, 1]. That means the actual skew is then 2 (calculated as 3 - 1), which violates maxSkew: 1. To satisfy the constraints and context for this example, the incoming Pod can only be placed onto a node in zone B:

OR

You can tweak the Pod spec to meet various kinds of requirements:

• Change maxSkew to a bigger value - such as 2 - so that the incoming Pod can be placed into zone A as well.
• Change topologyKey to node so as to distribute the Pods evenly across nodes instead of zones. In the above example, if maxSkew remains 1, the incoming Pod can only be placed onto the node node4.
• Change whenUnsatisfiable: DoNotSchedule to whenUnsatisfiable: ScheduleAnyway to ensure the incoming Pod to be always schedulable (suppose other scheduling APIs are satisfied). However, it's preferred to be placed into the topology domain which has fewer matching Pods. (Be aware that this preference is jointly normalized with other internal scheduling priorities such as resource usage ratio).

Example: multiple topology spread constraints

This builds upon the previous example. Suppose you have a 4-node cluster where 3 existing Pods labeled foo: bar are located on node1, node2 and node3 respectively:

You can combine two topology spread constraints to control the spread of Pods both by node and by zone:

pods/topology-spread-constraints/two-constraints.yaml

kind: Pod
apiVersion: v1
metadata:
  name: mypod
  labels:
    foo: bar
spec:
  topologySpreadConstraints:
  - maxSkew: 1
    topologyKey: zone
    whenUnsatisfiable: DoNotSchedule
    labelSelector:
      matchLabels:
        foo: bar
  - maxSkew: 1
    topologyKey: node
    whenUnsatisfiable: DoNotSchedule
    labelSelector:
      matchLabels:
        foo: bar
  containers:
  - name: pause
    image: registry.k8s.io/pause:3.1

In this case, to match the first constraint, the incoming Pod can only be placed onto nodes in zone B; while in terms of the second constraint, the incoming Pod can only be scheduled to the node node4. The scheduler only considers options that satisfy all defined constraints, so the only valid placement is onto node node4.

Example: conflicting topology spread constraints

Multiple constraints can lead to conflicts. Suppose you have a 3-node cluster across 2 zones:

If you were to apply two-constraints.yaml (the manifest from the previous example) to this cluster, you would see that the Pod mypod stays in the Pending state. This happens because: to satisfy the first constraint, the Pod mypod can only be placed into zone B; while in terms of the second constraint, the Pod mypod can only schedule to node node2. The intersection of the two constraints returns an empty set, and the scheduler cannot place the Pod.

To overcome this situation, you can either increase the value of maxSkew or modify one of the constraints to use whenUnsatisfiable: ScheduleAnyway. Depending on circumstances, you might also decide to delete an existing Pod manually - for example, if you are troubleshooting why a bug-fix rollout is not making progress.

Interaction with node affinity and node selectors

The scheduler will skip the non-matching nodes from the skew calculations if the incoming Pod has spec.nodeSelector or spec.affinity.nodeAffinity defined.

Example: topology spread constraints with node affinity

Suppose you have a 5-node cluster ranging across zones A to C:

and you know that zone C must be excluded. In this case, you can compose a manifest as below, so that Pod mypod will be placed into zone B instead of zone C. Similarly, Kubernetes also respects spec.nodeSelector.

pods/topology-spread-constraints/one-constraint-with-nodeaffinity.yaml

kind: Pod
apiVersion: v1
metadata:
  name: mypod
  labels:
    foo: bar
spec:
  topologySpreadConstraints:
  - maxSkew: 1
    topologyKey: zone
    whenUnsatisfiable: DoNotSchedule
    labelSelector:
      matchLabels:
        foo: bar
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: zone
            operator: NotIn
            values:
            - zoneC
  containers:
  - name: pause
    image: registry.k8s.io/pause:3.1

Implicit conventions

There are some implicit conventions worth noting here:

• Only the Pods holding the same namespace as the incoming Pod can be matching candidates.

• The scheduler only considers nodes that have all topologySpreadConstraints[*].topologyKey present at the same time. Nodes missing any of these topologyKeys are bypassed. This implies that:

  1. any Pods located on those bypassed nodes do not impact maxSkew calculation - in the above example, suppose the node node1 does not have a label "zone", then the 2 Pods will be disregarded, hence the incoming Pod will be scheduled into zone A.
  2. the incoming Pod has no chances to be scheduled onto this kind of nodes - in the above example, suppose a node node5 has the mistyped label zone-typo: zoneC (and no zone label set). After node node5 joins the cluster, it will be bypassed and Pods for this workload aren't scheduled there.

• Be aware of what will happen if the incoming Pod's topologySpreadConstraints[*].labelSelector doesn't match its own labels. In the above example, if you remove the incoming Pod's labels, it can still be placed onto nodes in zone B, since the constraints are still satisfied. However, after that placement, the degree of imbalance of the cluster remains unchanged - it's still zone A having 2 Pods labeled as foo: bar, and zone B having 1 Pod labeled as foo: bar. If this is not what you expect, update the workload's topologySpreadConstraints[*].labelSelector to match the labels in the pod template.

Cluster-level default constraints

It is possible to set default topology spread constraints for a cluster. Default topology spread constraints are applied to a Pod if, and only if:

• It doesn't define any constraints in its .spec.topologySpreadConstraints.
• It belongs to a Service, ReplicaSet, StatefulSet or ReplicationController.

Default constraints can be set as part of the PodTopologySpread plugin arguments in a scheduling profile. The constraints are specified with the same API above, except that labelSelector must be empty. The selectors are calculated from the Services, ReplicaSets, StatefulSets or ReplicationControllers that the Pod belongs to.

An example configuration might look like follows:

apiVersion: kubescheduler.config.k8s.io/v1beta3
kind: KubeSchedulerConfiguration

profiles:
  - schedulerName: default-scheduler
    pluginConfig:
      - name: PodTopologySpread
        args:
          defaultConstraints:
            - maxSkew: 1
              topologyKey: topology.kubernetes.io/zone
              whenUnsatisfiable: ScheduleAnyway
          defaultingType: List

Built-in default constraints

If you don't configure any cluster-level default constraints for pod topology spreading, then kube-scheduler acts as if you specified the following default topology constraints:

defaultConstraints:
  - maxSkew: 3
    topologyKey: "kubernetes.io/hostname"
    whenUnsatisfiable: ScheduleAnyway
  - maxSkew: 5
    topologyKey: "topology.kubernetes.io/zone"
    whenUnsatisfiable: ScheduleAnyway

Also, the legacy SelectorSpread plugin, which provides an equivalent behavior, is disabled by default.

If you don't want to use the default Pod spreading constraints for your cluster, you can disable those defaults by setting defaultingType to List and leaving empty defaultConstraints in the PodTopologySpread plugin configuration:

apiVersion: kubescheduler.config.k8s.io/v1beta3
kind: KubeSchedulerConfiguration

profiles:
  - schedulerName: default-scheduler
    pluginConfig:
      - name: PodTopologySpread
        args:
          defaultConstraints: []
          defaultingType: List

Comparison with podAffinity and podAntiAffinity

In Kubernetes, inter-Pod affinity and anti-affinity control how Pods are scheduled in relation to one another - either more packed or more scattered.

podAffinity

attracts Pods; you can try to pack any number of Pods into qualifying topology domain(s).

podAntiAffinity

repels Pods. If you set this to requiredDuringSchedulingIgnoredDuringExecution mode then only a single Pod can be scheduled into a single topology domain; if you choose preferredDuringSchedulingIgnoredDuringExecution then you lose the ability to enforce the constraint.

For finer control, you can specify topology spread constraints to distribute Pods across different topology domains - to achieve either high availability or cost-saving. This can also help on rolling update workloads and scaling out replicas smoothly.

For more context, see the Motivation section of the enhancement proposal about Pod topology spread constraints.

Known limitations

• There's no guarantee that the constraints remain satisfied when Pods are removed. For example, scaling down a Deployment may result in imbalanced Pods distribution.

  You can use a tool such as the Descheduler to rebalance the Pods distribution.

• Pods matched on tainted nodes are respected. See Issue 80921.

• The scheduler doesn't have prior knowledge of all the zones or other topology domains that a cluster has. They are determined from the existing nodes in the cluster. This could lead to a problem in autoscaled clusters, when a node pool (or node group) is scaled to zero nodes, and you're expecting the cluster to scale up, because, in this case, those topology domains won't be considered until there is at least one node in them.

  You can work around this by using a cluster autoscaling tool that is aware of Pod topology spread constraints and is also aware of the overall set of topology domains.

What's next

• The blog article Introducing PodTopologySpread explains maxSkew in some detail, as well as covering some advanced usage examples.
• Read the scheduling section of the API reference for Pod.

6 - Taints and Tolerations

Node affinity is a property of Pods that attracts them to a set of nodes (either as a preference or a hard requirement). Taints are the opposite -- they allow a node to repel a set of pods.

Tolerations are applied to pods. Tolerations allow the scheduler to schedule pods with matching taints. Tolerations allow scheduling but don't guarantee scheduling: the scheduler also evaluates other parameters as part of its function.

Taints and tolerations work together to ensure that pods are not scheduled onto inappropriate nodes. One or more taints are applied to a node; this marks that the node should not accept any pods that do not tolerate the taints.

Concepts

You add a taint to a node using kubectl taint. For example,

kubectl taint nodes node1 key1=value1:NoSchedule

places a taint on node node1. The taint has key key1, value value1, and taint effect NoSchedule. This means that no pod will be able to schedule onto node1 unless it has a matching toleration.

To remove the taint added by the command above, you can run:

kubectl taint nodes node1 key1=value1:NoSchedule-

You specify a toleration for a pod in the PodSpec. Both of the following tolerations "match" the taint created by the kubectl taint line above, and thus a pod with either toleration would be able to schedule onto node1:

tolerations:
- key: "key1"
  operator: "Equal"
  value: "value1"
  effect: "NoSchedule"

tolerations:
- key: "key1"
  operator: "Exists"
  effect: "NoSchedule"

The default Kubernetes scheduler takes taints and tolerations into account when selecting a node to run a particular Pod. However, if you manually specify the .spec.nodeName for a Pod, that action bypasses the scheduler; the Pod is then bound onto the node where you assigned it, even if there are NoSchedule taints on that node that you selected. If this happens and the node also has a NoExecute taint set, the kubelet will eject the Pod unless there is an appropriate tolerance set.

Here's an example of a pod that has some tolerations defined:

pods/pod-with-toleration.yaml

apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    env: test
spec:
  containers:
  - name: nginx
    image: nginx
    imagePullPolicy: IfNotPresent
  tolerations:
  - key: "example-key"
    operator: "Exists"
    effect: "NoSchedule"

The default value for operator is Equal.

A toleration "matches" a taint if the keys are the same and the effects are the same, and:

• the operator is Exists (in which case no value should be specified), or
• the operator is Equal and the values should be equal.

The above example used the effect of NoSchedule. Alternatively, you can use the effect of PreferNoSchedule.

The allowed values for the effect field are:

NoExecute

This affects pods that are already running on the node as follows:

• Pods that do not tolerate the taint are evicted immediately
• Pods that tolerate the taint without specifying tolerationSeconds in their toleration specification remain bound forever
• Pods that tolerate the taint with a specified tolerationSeconds remain bound for the specified amount of time. After that time elapses, the node lifecycle controller evicts the Pods from the node.

NoSchedule

No new Pods will be scheduled on the tainted node unless they have a matching toleration. Pods currently running on the node are not evicted.

PreferNoSchedule

PreferNoSchedule is a "preference" or "soft" version of NoSchedule. The control plane will try to avoid placing a Pod that does not tolerate the taint on the node, but it is not guaranteed.

You can put multiple taints on the same node and multiple tolerations on the same pod. The way Kubernetes processes multiple taints and tolerations is like a filter: start with all of a node's taints, then ignore the ones for which the pod has a matching toleration; the remaining un-ignored taints have the indicated effects on the pod. In particular,

• if there is at least one un-ignored taint with effect NoSchedule then Kubernetes will not schedule the pod onto that node
• if there is no un-ignored taint with effect NoSchedule but there is at least one un-ignored taint with effect PreferNoSchedule then Kubernetes will try to not schedule the pod onto the node
• if there is at least one un-ignored taint with effect NoExecute then the pod will be evicted from the node (if it is already running on the node), and will not be scheduled onto the node (if it is not yet running on the node).

For example, imagine you taint a node like this

kubectl taint nodes node1 key1=value1:NoSchedule
kubectl taint nodes node1 key1=value1:NoExecute
kubectl taint nodes node1 key2=value2:NoSchedule

And a pod has two tolerations:

tolerations:
- key: "key1"
  operator: "Equal"
  value: "value1"
  effect: "NoSchedule"
- key: "key1"
  operator: "Equal"
  value: "value1"
  effect: "NoExecute"

In this case, the pod will not be able to schedule onto the node, because there is no toleration matching the third taint. But it will be able to continue running if it is already running on the node when the taint is added, because the third taint is the only one of the three that is not tolerated by the pod.

Normally, if a taint with effect NoExecute is added to a node, then any pods that do not tolerate the taint will be evicted immediately, and pods that do tolerate the taint will never be evicted. However, a toleration with NoExecute effect can specify an optional tolerationSeconds field that dictates how long the pod will stay bound to the node after the taint is added. For example,

tolerations:
- key: "key1"
  operator: "Equal"
  value: "value1"
  effect: "NoExecute"
  tolerationSeconds: 3600

means that if this pod is running and a matching taint is added to the node, then the pod will stay bound to the node for 3600 seconds, and then be evicted. If the taint is removed before that time, the pod will not be evicted.

Example Use Cases

Taints and tolerations are a flexible way to steer pods away from nodes or evict pods that shouldn't be running. A few of the use cases are

• Dedicated Nodes: If you want to dedicate a set of nodes for exclusive use by a particular set of users, you can add a taint to those nodes (say, kubectl taint nodes nodename dedicated=groupName:NoSchedule) and then add a corresponding toleration to their pods (this would be done most easily by writing a custom admission controller). The pods with the tolerations will then be allowed to use the tainted (dedicated) nodes as well as any other nodes in the cluster. If you want to dedicate the nodes to them and ensure they only use the dedicated nodes, then you should additionally add a label similar to the taint to the same set of nodes (e.g. dedicated=groupName), and the admission controller should additionally add a node affinity to require that the pods can only schedule onto nodes labeled with dedicated=groupName.

• Nodes with Special Hardware: In a cluster where a small subset of nodes have specialized hardware (for example GPUs), it is desirable to keep pods that don't need the specialized hardware off of those nodes, thus leaving room for later-arriving pods that do need the specialized hardware. This can be done by tainting the nodes that have the specialized hardware (e.g. kubectl taint nodes nodename special=true:NoSchedule or kubectl taint nodes nodename special=true:PreferNoSchedule) and adding a corresponding toleration to pods that use the special hardware. As in the dedicated nodes use case, it is probably easiest to apply the tolerations using a custom admission controller. For example, it is recommended to use Extended Resources to represent the special hardware, taint your special hardware nodes with the extended resource name and run the ExtendedResourceToleration admission controller. Now, because the nodes are tainted, no pods without the toleration will schedule on them. But when you submit a pod that requests the extended resource, the ExtendedResourceToleration admission controller will automatically add the correct toleration to the pod and that pod will schedule on the special hardware nodes. This will make sure that these special hardware nodes are dedicated for pods requesting such hardware and you don't have to manually add tolerations to your pods.

• Taint based Evictions: A per-pod-configurable eviction behavior when there are node problems, which is described in the next section.

Taint based Evictions

The node controller automatically taints a Node when certain conditions are true. The following taints are built in:

• node.kubernetes.io/not-ready: Node is not ready. This corresponds to the NodeCondition Ready being "False".
• node.kubernetes.io/unreachable: Node is unreachable from the node controller. This corresponds to the NodeCondition Ready being "Unknown".
• node.kubernetes.io/memory-pressure: Node has memory pressure.
• node.kubernetes.io/disk-pressure: Node has disk pressure.
• node.kubernetes.io/pid-pressure: Node has PID pressure.
• node.kubernetes.io/network-unavailable: Node's network is unavailable.
• node.kubernetes.io/unschedulable: Node is unschedulable.
• node.cloudprovider.kubernetes.io/uninitialized: When the kubelet is started with an "external" cloud provider, this taint is set on a node to mark it as unusable. After a controller from the cloud-controller-manager initializes this node, the kubelet removes this taint.

In case a node is to be drained, the node controller or the kubelet adds relevant taints with NoExecute effect. This effect is added by default for the node.kubernetes.io/not-ready and node.kubernetes.io/unreachable taints. If the fault condition returns to normal, the kubelet or node controller can remove the relevant taint(s).

In some cases when the node is unreachable, the API server is unable to communicate with the kubelet on the node. The decision to delete the pods cannot be communicated to the kubelet until communication with the API server is re-established. In the meantime, the pods that are scheduled for deletion may continue to run on the partitioned node.

You can specify tolerationSeconds for a Pod to define how long that Pod stays bound to a failing or unresponsive Node.

For example, you might want to keep an application with a lot of local state bound to node for a long time in the event of network partition, hoping that the partition will recover and thus the pod eviction can be avoided. The toleration you set for that Pod might look like:

tolerations:
- key: "node.kubernetes.io/unreachable"
  operator: "Exists"
  effect: "NoExecute"
  tolerationSeconds: 6000

DaemonSet pods are created with NoExecute tolerations for the following taints with no tolerationSeconds:

• node.kubernetes.io/unreachable
• node.kubernetes.io/not-ready

This ensures that DaemonSet pods are never evicted due to these problems.

Taint Nodes by Condition

The control plane, using the node controller, automatically creates taints with a NoSchedule effect for node conditions.

The scheduler checks taints, not node conditions, when it makes scheduling decisions. This ensures that node conditions don't directly affect scheduling. For example, if the DiskPressure node condition is active, the control plane adds the node.kubernetes.io/disk-pressure taint and does not schedule new pods onto the affected node. If the MemoryPressure node condition is active, the control plane adds the node.kubernetes.io/memory-pressure taint.

You can ignore node conditions for newly created pods by adding the corresponding Pod tolerations. The control plane also adds the node.kubernetes.io/memory-pressure toleration on pods that have a QoS class other than BestEffort. This is because Kubernetes treats pods in the Guaranteed or Burstable QoS classes (even pods with no memory request set) as if they are able to cope with memory pressure, while new BestEffort pods are not scheduled onto the affected node.

The DaemonSet controller automatically adds the following NoSchedule tolerations to all daemons, to prevent DaemonSets from breaking.

• node.kubernetes.io/memory-pressure
• node.kubernetes.io/disk-pressure
• node.kubernetes.io/pid-pressure (1.14 or later)
• node.kubernetes.io/unschedulable (1.10 or later)
• node.kubernetes.io/network-unavailable (host network only)

Adding these tolerations ensures backward compatibility. You can also add arbitrary tolerations to DaemonSets.

What's next

• Read about Node-pressure Eviction and how you can configure it
• Read about Pod Priority

7 - Scheduling Framework

The scheduling framework is a pluggable architecture for the Kubernetes scheduler. It consists of a set of "plugin" APIs that are compiled directly into the scheduler. These APIs allow most scheduling features to be implemented as plugins, while keeping the scheduling "core" lightweight and maintainable. Refer to the design proposal of the scheduling framework for more technical information on the design of the framework.

Framework workflow

The Scheduling Framework defines a few extension points. Scheduler plugins register to be invoked at one or more extension points. Some of these plugins can change the scheduling decisions and some are informational only.

Each attempt to schedule one Pod is split into two phases, the scheduling cycle and the binding cycle.

Scheduling cycle & binding cycle

The scheduling cycle selects a node for the Pod, and the binding cycle applies that decision to the cluster. Together, a scheduling cycle and binding cycle are referred to as a "scheduling context".

Scheduling cycles are run serially, while binding cycles may run concurrently.

A scheduling or binding cycle can be aborted if the Pod is determined to be unschedulable or if there is an internal error. The Pod will be returned to the queue and retried.

Interfaces

The following picture shows the scheduling context of a Pod and the interfaces that the scheduling framework exposes.

One plugin may implement multiple interfaces to perform more complex or stateful tasks.

Some interfaces match the scheduler extension points which can be configured through Scheduler Configuration.

PreEnqueue

These plugins are called prior to adding Pods to the internal active queue, where Pods are marked as ready for scheduling.

Only when all PreEnqueue plugins return Success, the Pod is allowed to enter the active queue. Otherwise, it's placed in the internal unschedulable Pods list, and doesn't get an Unschedulable condition.

For more details about how internal scheduler queues work, read Scheduling queue in kube-scheduler.

EnqueueExtension

EnqueueExtension is the interface where the plugin can control whether to retry scheduling of Pods rejected by this plugin, based on changes in the cluster. Plugins that implement PreEnqueue, PreFilter, Filter, Reserve or Permit should implement this interface.

QueueingHint

QueueingHint is a callback function for deciding whether a Pod can be requeued to the active queue or backoff queue. It's executed every time a certain kind of event or change happens in the cluster. When the QueueingHint finds that the event might make the Pod schedulable, the Pod is put into the active queue or the backoff queue so that the scheduler will retry the scheduling of the Pod.

QueueSort

These plugins are used to sort Pods in the scheduling queue. A queue sort plugin essentially provides a Less(Pod1, Pod2) function. Only one queue sort plugin may be enabled at a time.

PreFilter

These plugins are used to pre-process info about the Pod, or to check certain conditions that the cluster or the Pod must meet. If a PreFilter plugin returns an error, the scheduling cycle is aborted.

Filter

These plugins are used to filter out nodes that cannot run the Pod. For each node, the scheduler will call filter plugins in their configured order. If any filter plugin marks the node as infeasible, the remaining plugins will not be called for that node. Nodes may be evaluated concurrently.

PostFilter

These plugins are called after the Filter phase, but only when no feasible nodes were found for the pod. Plugins are called in their configured order. If any postFilter plugin marks the node as Schedulable, the remaining plugins will not be called. A typical PostFilter implementation is preemption, which tries to make the pod schedulable by preempting other Pods.

PreScore

These plugins are used to perform "pre-scoring" work, which generates a sharable state for Score plugins to use. If a PreScore plugin returns an error, the scheduling cycle is aborted.

Score

These plugins are used to rank nodes that have passed the filtering phase. The scheduler will call each scoring plugin for each node. There will be a well defined range of integers representing the minimum and maximum scores. After the NormalizeScore phase, the scheduler will combine node scores from all plugins according to the configured plugin weights.

NormalizeScore

These plugins are used to modify scores before the scheduler computes a final ranking of Nodes. A plugin that registers for this extension point will be called with the Score results from the same plugin. This is called once per plugin per scheduling cycle.

For example, suppose a plugin BlinkingLightScorer ranks Nodes based on how many blinking lights they have.

func ScoreNode(_ *v1.pod, n *v1.Node) (int, error) {
    return getBlinkingLightCount(n)
}

However, the maximum count of blinking lights may be small compared to NodeScoreMax. To fix this, BlinkingLightScorer should also register for this extension point.

func NormalizeScores(scores map[string]int) {
    highest := 0
    for _, score := range scores {
        highest = max(highest, score)
    }
    for node, score := range scores {
        scores[node] = score*NodeScoreMax/highest
    }
}

If any NormalizeScore plugin returns an error, the scheduling cycle is aborted.

Reserve

A plugin that implements the Reserve interface has two methods, namely Reserve and Unreserve, that back two informational scheduling phases called Reserve and Unreserve, respectively. Plugins which maintain runtime state (aka "stateful plugins") should use these phases to be notified by the scheduler when resources on a node are being reserved and unreserved for a given Pod.

The Reserve phase happens before the scheduler actually binds a Pod to its designated node. It exists to prevent race conditions while the scheduler waits for the bind to succeed. The Reserve method of each Reserve plugin may succeed or fail; if one Reserve method call fails, subsequent plugins are not executed and the Reserve phase is considered to have failed. If the Reserve method of all plugins succeed, the Reserve phase is considered to be successful and the rest of the scheduling cycle and the binding cycle are executed.

The Unreserve phase is triggered if the Reserve phase or a later phase fails. When this happens, the Unreserve method of all Reserve plugins will be executed in the reverse order of Reserve method calls. This phase exists to clean up the state associated with the reserved Pod.

Permit

Permit plugins are invoked at the end of the scheduling cycle for each Pod, to prevent or delay the binding to the candidate node. A permit plugin can do one of the three things:

1. approve
   Once all Permit plugins approve a Pod, it is sent for binding.

2. deny
   If any Permit plugin denies a Pod, it is returned to the scheduling queue. This will trigger the Unreserve phase in Reserve plugins.

3. wait (with a timeout)
   If a Permit plugin returns "wait", then the Pod is kept in an internal "waiting" Pods list, and the binding cycle of this Pod starts but directly blocks until it gets approved. If a timeout occurs, wait becomes deny and the Pod is returned to the scheduling queue, triggering the Unreserve phase in Reserve plugins.

PreBind

These plugins are used to perform any work required before a Pod is bound. For example, a pre-bind plugin may provision a network volume and mount it on the target node before allowing the Pod to run there.

If any PreBind plugin returns an error, the Pod is rejected and returned to the scheduling queue.

Bind

These plugins are used to bind a Pod to a Node. Bind plugins will not be called until all PreBind plugins have completed. Each bind plugin is called in the configured order. A bind plugin may choose whether or not to handle the given Pod. If a bind plugin chooses to handle a Pod, the remaining bind plugins are skipped.

PostBind

This is an informational interface. Post-bind plugins are called after a Pod is successfully bound. This is the end of a binding cycle, and can be used to clean up associated resources.

Plugin API

There are two steps to the plugin API. First, plugins must register and get configured, then they use the extension point interfaces. Extension point interfaces have the following form.

type Plugin interface {
    Name() string
}

type QueueSortPlugin interface {
    Plugin
    Less(*v1.pod, *v1.pod) bool
}

type PreFilterPlugin interface {
    Plugin
    PreFilter(context.Context, *framework.CycleState, *v1.pod) error
}

// ...

Plugin configuration

You can enable or disable plugins in the scheduler configuration. If you are using Kubernetes v1.18 or later, most scheduling plugins are in use and enabled by default.

In addition to default plugins, you can also implement your own scheduling plugins and get them configured along with default plugins. You can visit scheduler-plugins for more details.

If you are using Kubernetes v1.18 or later, you can configure a set of plugins as a scheduler profile and then define multiple profiles to fit various kinds of workload. Learn more at multiple profiles.

8 - Dynamic Resource Allocation

Dynamic resource allocation is an API for requesting and sharing resources between pods and containers inside a pod. It is a generalization of the persistent volumes API for generic resources. Typically those resources are devices like GPUs.

Third-party resource drivers are responsible for tracking and preparing resources, with allocation of resources handled by Kubernetes via structured parameters (introduced in Kubernetes 1.30). Different kinds of resources support arbitrary parameters for defining requirements and initialization.

Kubernetes v1.26 through to 1.31 included an (alpha) implementation of classic DRA, which is no longer supported. This documentation, which is for Kubernetes v1.32, explains the current approach to dynamic resource allocation within Kubernetes.

Before you begin

Kubernetes v1.32 includes cluster-level API support for dynamic resource allocation, but it needs to be enabled explicitly. You also must install a resource driver for specific resources that are meant to be managed using this API. If you are not running Kubernetes v1.32, check the documentation for that version of Kubernetes.

API

The resource.k8s.io/v1beta1 API group provides these types:

ResourceClaim

Describes a request for access to resources in the cluster, for use by workloads. For example, if a workload needs an accelerator device with specific properties, this is how that request is expressed. The status stanza tracks whether this claim has been satisfied and what specific resources have been allocated.

ResourceClaimTemplate

Defines the spec and some metadata for creating ResourceClaims. Created by a user when deploying a workload. The per-Pod ResourceClaims are then created and removed by Kubernetes automatically.

DeviceClass

Contains pre-defined selection criteria for certain devices and configuration for them. DeviceClasses are created by a cluster administrator when installing a resource driver. Each request to allocate a device in a ResourceClaim must reference exactly one DeviceClass.

ResourceSlice

Used by DRA drivers to publish information about resources that are available in the cluster.

All parameters that select devices are defined in the ResourceClaim and DeviceClass with in-tree types. Configuration parameters can be embedded there. Which configuration parameters are valid depends on the DRA driver -- Kubernetes only passes them through without interpreting them.

The core/v1 PodSpec defines ResourceClaims that are needed for a Pod in a resourceClaims field. Entries in that list reference either a ResourceClaim or a ResourceClaimTemplate. When referencing a ResourceClaim, all Pods using this PodSpec (for example, inside a Deployment or StatefulSet) share the same ResourceClaim instance. When referencing a ResourceClaimTemplate, each Pod gets its own instance.

The resources.claims list for container resources defines whether a container gets access to these resource instances, which makes it possible to share resources between one or more containers.

Here is an example for a fictional resource driver. Two ResourceClaim objects will get created for this Pod and each container gets access to one of them.

apiVersion: resource.k8s.io/v1beta1
kind: DeviceClass
name: resource.example.com
spec:
  selectors:
  - cel:
      expression: device.driver == "resource-driver.example.com"
---
apiVersion: resource.k8s.io/v1beta1
kind: ResourceClaimTemplate
metadata:
  name: large-black-cat-claim-template
spec:
  spec:
    devices:
      requests:
      - name: req-0
        deviceClassName: resource.example.com
        selectors:
        - cel:
           expression: |-
              device.attributes["resource-driver.example.com"].color == "black" &&
              device.attributes["resource-driver.example.com"].size == "large"              
–--
apiVersion: v1
kind: Pod
metadata:
  name: pod-with-cats
spec:
  containers:
  - name: container0
    image: ubuntu:20.04
    command: ["sleep", "9999"]
    resources:
      claims:
      - name: cat-0
  - name: container1
    image: ubuntu:20.04
    command: ["sleep", "9999"]
    resources:
      claims:
      - name: cat-1
  resourceClaims:
  - name: cat-0
    resourceClaimTemplateName: large-black-cat-claim-template
  - name: cat-1
    resourceClaimTemplateName: large-black-cat-claim-template

Scheduling

The scheduler is responsible for allocating resources to a ResourceClaim whenever a pod needs them. It does so by retrieving the full list of available resources from ResourceSlice objects, tracking which of those resources have already been allocated to existing ResourceClaims, and then selecting from those resources that remain.

The only kind of supported resources at the moment are devices. A device instance has a name and several attributes and capacities. Devices get selected through CEL expressions which check those attributes and capacities. In addition, the set of selected devices also can be restricted to sets which meet certain constraints.

The chosen resource is recorded in the ResourceClaim status together with any vendor-specific configuration, so when a pod is about to start on a node, the resource driver on the node has all the information it needs to prepare the resource.

By using structured parameters, the scheduler is able to reach a decision without communicating with any DRA resource drivers. It is also able to schedule multiple pods quickly by keeping information about ResourceClaim allocations in memory and writing this information to the ResourceClaim objects in the background while concurrently binding the pod to a node.

Monitoring resources

The kubelet provides a gRPC service to enable discovery of dynamic resources of running Pods. For more information on the gRPC endpoints, see the resource allocation reporting.

Pre-scheduled Pods

When you - or another API client - create a Pod with spec.nodeName already set, the scheduler gets bypassed. If some ResourceClaim needed by that Pod does not exist yet, is not allocated or not reserved for the Pod, then the kubelet will fail to run the Pod and re-check periodically because those requirements might still get fulfilled later.

Such a situation can also arise when support for dynamic resource allocation was not enabled in the scheduler at the time when the Pod got scheduled (version skew, configuration, feature gate, etc.). kube-controller-manager detects this and tries to make the Pod runnable by reserving the required ResourceClaims. However, this only works if those were allocated by the scheduler for some other pod.

It is better to avoid bypassing the scheduler because a Pod that is assigned to a node blocks normal resources (RAM, CPU) that then cannot be used for other Pods while the Pod is stuck. To make a Pod run on a specific node while still going through the normal scheduling flow, create the Pod with a node selector that exactly matches the desired node:

apiVersion: v1
kind: Pod
metadata:
  name: pod-with-cats
spec:
  nodeSelector:
    kubernetes.io/hostname: name-of-the-intended-node
  ...

You may also be able to mutate the incoming Pod, at admission time, to unset the .spec.nodeName field and to use a node selector instead.

Admin access

You can mark a request in a ResourceClaim or ResourceClaimTemplate as having privileged features. A request with admin access grants access to devices which are in use and may enable additional permissions when making the device available in a container:

apiVersion: resource.k8s.io/v1beta1
kind: ResourceClaimTemplate
metadata:
  name: large-black-cat-claim-template
spec:
  spec:
    devices:
      requests:
      - name: req-0
        deviceClassName: resource.example.com
        adminAccess: true

If this feature is disabled, the adminAccess field will be removed automatically when creating such a ResourceClaim.

Admin access is a privileged mode which should not be made available to normal users in a multi-tenant cluster. Cluster administrators can restrict usage of this feature by installing a validating admission policy similar to the following example. Cluster administrators need to adapt at least the names and replace "dra.example.com".

# Permission to use admin access is granted only in namespaces which have the
# "admin-access.dra.example.com" label. Other ways of making that decision are
# also possible.

apiVersion: admissionregistration.k8s.io/v1
kind: ValidatingAdmissionPolicy
metadata:
  name: resourceclaim-policy.dra.example.com
spec:
  failurePolicy: Fail
  matchConstraints:
    resourceRules:
    - apiGroups:   ["resource.k8s.io"]
      apiVersions: ["v1alpha3", "v1beta1"]
      operations:  ["CREATE", "UPDATE"]
      resources:   ["resourceclaims"]
  validations:
    - expression: '! object.spec.devices.requests.exists(e, has(e.adminAccess) && e.adminAccess)'
      reason: Forbidden
      messageExpression: '"admin access to devices not enabled"'
---
apiVersion: admissionregistration.k8s.io/v1
kind: ValidatingAdmissionPolicyBinding
metadata:
  name: resourceclaim-binding.dra.example.com
spec:
  policyName:  resourceclaim-policy.dra.example.com
  validationActions: [Deny]
  matchResources:
    namespaceSelector:
      matchExpressions:
      - key: admin-access.dra.example.com
        operator: DoesNotExist
---
apiVersion: admissionregistration.k8s.io/v1
kind: ValidatingAdmissionPolicy
metadata:
  name: resourceclaimtemplate-policy.dra.example.com
spec:
  failurePolicy: Fail
  matchConstraints:
    resourceRules:
    - apiGroups:   ["resource.k8s.io"]
      apiVersions: ["v1alpha3", "v1beta1"]
      operations:  ["CREATE", "UPDATE"]
      resources:   ["resourceclaimtemplates"]
  validations:
    - expression: '! object.spec.spec.devices.requests.exists(e, has(e.adminAccess) && e.adminAccess)'
      reason: Forbidden
      messageExpression: '"admin access to devices not enabled"'
---
apiVersion: admissionregistration.k8s.io/v1
kind: ValidatingAdmissionPolicyBinding
metadata:
  name: resourceclaimtemplate-binding.dra.example.com
spec:
  policyName:  resourceclaimtemplate-policy.dra.example.com
  validationActions: [Deny]
  matchResources:
    namespaceSelector:
      matchExpressions:
      - key: admin-access.dra.example.com
        operator: DoesNotExist

ResourceClaim Device Status

The drivers can report driver-specific device status data for each allocated device in a resource claim. For example, IPs assigned to a network interface device can be reported in the ResourceClaim status.

The drivers setting the status, the accuracy of the information depends on the implementation of those DRA Drivers. Therefore, the reported status of the device may not always reflect the real time changes of the state of the device.

When the feature is disabled, that field automatically gets cleared when storing the ResourceClaim.

A ResourceClaim device status is supported when it is possible, from a DRA driver, to update an existing ResourceClaim where the status.devices field is set.

Enabling dynamic resource allocation

Dynamic resource allocation is a beta feature which is off by default and only enabled when the DynamicResourceAllocation feature gate and the resource.k8s.io/v1beta1 API group are enabled. For details on that, see the --feature-gates and --runtime-config kube-apiserver parameters. kube-scheduler, kube-controller-manager and kubelet also need the feature gate.

When a resource driver reports the status of the devices, then the DRAResourceClaimDeviceStatus feature gate has to be enabled in addition to DynamicResourceAllocation.

A quick check whether a Kubernetes cluster supports the feature is to list DeviceClass objects with:

kubectl get deviceclasses

If your cluster supports dynamic resource allocation, the response is either a list of DeviceClass objects or:

No resources found

If not supported, this error is printed instead:

error: the server doesn't have a resource type "deviceclasses"

The default configuration of kube-scheduler enables the "DynamicResources" plugin if and only if the feature gate is enabled and when using the v1 configuration API. Custom configurations may have to be modified to include it.

In addition to enabling the feature in the cluster, a resource driver also has to be installed. Please refer to the driver's documentation for details.

Enabling admin access

Admin access is an alpha feature and only enabled when the DRAAdminAccess feature gate is enabled in the kube-apiserver and kube-scheduler.

Enabling Device Status

ResourceClaim Device Status is an alpha feature and only enabled when the DRAResourceClaimDeviceStatus feature gate is enabled in the kube-apiserver.

What's next

• For more information on the design, see the Dynamic Resource Allocation with Structured Parameters KEP.

9 - Scheduler Performance Tuning

kube-scheduler is the Kubernetes default scheduler. It is responsible for placement of Pods on Nodes in a cluster.

Nodes in a cluster that meet the scheduling requirements of a Pod are called feasible Nodes for the Pod. The scheduler finds feasible Nodes for a Pod and then runs a set of functions to score the feasible Nodes, picking a Node with the highest score among the feasible ones to run the Pod. The scheduler then notifies the API server about this decision in a process called Binding.

This page explains performance tuning optimizations that are relevant for large Kubernetes clusters.

In large clusters, you can tune the scheduler's behaviour balancing scheduling outcomes between latency (new Pods are placed quickly) and accuracy (the scheduler rarely makes poor placement decisions).

You configure this tuning setting via kube-scheduler setting percentageOfNodesToScore. This KubeSchedulerConfiguration setting determines a threshold for scheduling nodes in your cluster.

Setting the threshold

The percentageOfNodesToScore option accepts whole numeric values between 0 and 100. The value 0 is a special number which indicates that the kube-scheduler should use its compiled-in default. If you set percentageOfNodesToScore above 100, kube-scheduler acts as if you had set a value of 100.

To change the value, edit the kube-scheduler configuration file and then restart the scheduler. In many cases, the configuration file can be found at /etc/kubernetes/config/kube-scheduler.yaml.

After you have made this change, you can run

kubectl get pods -n kube-system | grep kube-scheduler

to verify that the kube-scheduler component is healthy.

Node scoring threshold

To improve scheduling performance, the kube-scheduler can stop looking for feasible nodes once it has found enough of them. In large clusters, this saves time compared to a naive approach that would consider every node.

You specify a threshold for how many nodes are enough, as a whole number percentage of all the nodes in your cluster. The kube-scheduler converts this into an integer number of nodes. During scheduling, if the kube-scheduler has identified enough feasible nodes to exceed the configured percentage, the kube-scheduler stops searching for more feasible nodes and moves on to the scoring phase.

How the scheduler iterates over Nodes describes the process in detail.

Default threshold

If you don't specify a threshold, Kubernetes calculates a figure using a linear formula that yields 50% for a 100-node cluster and yields 10% for a 5000-node cluster. The lower bound for the automatic value is 5%.

This means that the kube-scheduler always scores at least 5% of your cluster no matter how large the cluster is, unless you have explicitly set percentageOfNodesToScore to be smaller than 5.

If you want the scheduler to score all nodes in your cluster, set percentageOfNodesToScore to 100.

Example

Below is an example configuration that sets percentageOfNodesToScore to 50%.

apiVersion: kubescheduler.config.k8s.io/v1alpha1
kind: KubeSchedulerConfiguration
algorithmSource:
  provider: DefaultProvider

...

percentageOfNodesToScore: 50

Tuning percentageOfNodesToScore

percentageOfNodesToScore must be a value between 1 and 100 with the default value being calculated based on the cluster size. There is also a hardcoded minimum value of 100 nodes.

An important detail to consider when setting this value is that when a smaller number of nodes in a cluster are checked for feasibility, some nodes are not sent to be scored for a given Pod. As a result, a Node which could possibly score a higher value for running the given Pod might not even be passed to the scoring phase. This would result in a less than ideal placement of the Pod.

You should avoid setting percentageOfNodesToScore very low so that kube-scheduler does not make frequent, poor Pod placement decisions. Avoid setting the percentage to anything below 10%, unless the scheduler's throughput is critical for your application and the score of nodes is not important. In other words, you prefer to run the Pod on any Node as long as it is feasible.

How the scheduler iterates over Nodes

This section is intended for those who want to understand the internal details of this feature.

In order to give all the Nodes in a cluster a fair chance of being considered for running Pods, the scheduler iterates over the nodes in a round robin fashion. You can imagine that Nodes are in an array. The scheduler starts from the start of the array and checks feasibility of the nodes until it finds enough Nodes as specified by percentageOfNodesToScore. For the next Pod, the scheduler continues from the point in the Node array that it stopped at when checking feasibility of Nodes for the previous Pod.

If Nodes are in multiple zones, the scheduler iterates over Nodes in various zones to ensure that Nodes from different zones are considered in the feasibility checks. As an example, consider six nodes in two zones:

Zone 1: Node 1, Node 2, Node 3, Node 4
Zone 2: Node 5, Node 6

The Scheduler evaluates feasibility of the nodes in this order:

Node 1, Node 5, Node 2, Node 6, Node 3, Node 4

After going over all the Nodes, it goes back to Node 1.

What's next

• Check the kube-scheduler configuration reference (v1)

10 - Resource Bin Packing

In the scheduling-plugin NodeResourcesFit of kube-scheduler, there are two scoring strategies that support the bin packing of resources: MostAllocated and RequestedToCapacityRatio.

Enabling bin packing using MostAllocated strategy

The MostAllocated strategy scores the nodes based on the utilization of resources, favoring the ones with higher allocation. For each resource type, you can set a weight to modify its influence in the node score.

To set the MostAllocated strategy for the NodeResourcesFit plugin, use a scheduler configuration similar to the following:

apiVersion: kubescheduler.config.k8s.io/v1
kind: KubeSchedulerConfiguration
profiles:
- pluginConfig:
  - args:
      scoringStrategy:
        resources:
        - name: cpu
          weight: 1
        - name: memory
          weight: 1
        - name: intel.com/foo
          weight: 3
        - name: intel.com/bar
          weight: 3
        type: MostAllocated
    name: NodeResourcesFit

To learn more about other parameters and their default configuration, see the API documentation for NodeResourcesFitArgs.

Enabling bin packing using RequestedToCapacityRatio

The RequestedToCapacityRatio strategy allows the users to specify the resources along with weights for each resource to score nodes based on the request to capacity ratio. This allows users to bin pack extended resources by using appropriate parameters to improve the utilization of scarce resources in large clusters. It favors nodes according to a configured function of the allocated resources. The behavior of the RequestedToCapacityRatio in the NodeResourcesFit score function can be controlled by the scoringStrategy field. Within the scoringStrategy field, you can configure two parameters: requestedToCapacityRatio and resources. The shape in the requestedToCapacityRatio parameter allows the user to tune the function as least requested or most requested based on utilization and score values. The resources parameter comprises both the name of the resource to be considered during scoring and its corresponding weight, which specifies the weight of each resource.

Below is an example configuration that sets the bin packing behavior for extended resources intel.com/foo and intel.com/bar using the requestedToCapacityRatio field.

apiVersion: kubescheduler.config.k8s.io/v1
kind: KubeSchedulerConfiguration
profiles:
- pluginConfig:
  - args:
      scoringStrategy:
        resources:
        - name: intel.com/foo
          weight: 3
        - name: intel.com/bar
          weight: 3
        requestedToCapacityRatio:
          shape:
          - utilization: 0
            score: 0
          - utilization: 100
            score: 10
        type: RequestedToCapacityRatio
    name: NodeResourcesFit

Referencing the KubeSchedulerConfiguration file with the kube-scheduler flag --config=/path/to/config/file will pass the configuration to the scheduler.

To learn more about other parameters and their default configuration, see the API documentation for NodeResourcesFitArgs.

Tuning the score function

shape is used to specify the behavior of the RequestedToCapacityRatio function.

shape:
  - utilization: 0
    score: 0
  - utilization: 100
    score: 10

The above arguments give the node a score of 0 if utilization is 0% and 10 for utilization 100%, thus enabling bin packing behavior. To enable least requested the score value must be reversed as follows.

shape:
  - utilization: 0
    score: 10
  - utilization: 100
    score: 0

resources is an optional parameter which defaults to:

resources:
  - name: cpu
    weight: 1
  - name: memory
    weight: 1

It can be used to add extended resources as follows:

resources:
  - name: intel.com/foo
    weight: 5
  - name: cpu
    weight: 3
  - name: memory
    weight: 1

The weight parameter is optional and is set to 1 if not specified. Also, the weight cannot be set to a negative value.

Node scoring for capacity allocation

This section is intended for those who want to understand the internal details of this feature. Below is an example of how the node score is calculated for a given set of values.

Requested resources:

intel.com/foo : 2
memory: 256MB
cpu: 2

Resource weights:

intel.com/foo : 5
memory: 1
cpu: 3

FunctionShapePoint {{0, 0}, {100, 10}}

Node 1 spec:

Available:
  intel.com/foo: 4
  memory: 1 GB
  cpu: 8

Used:
  intel.com/foo: 1
  memory: 256MB
  cpu: 1

Node score:

intel.com/foo  = resourceScoringFunction((2+1),4)
               = (100 - ((4-3)*100/4)
               = (100 - 25)
               = 75                       # requested + used = 75% * available
               = rawScoringFunction(75)
               = 7                        # floor(75/10)

memory         = resourceScoringFunction((256+256),1024)
               = (100 -((1024-512)*100/1024))
               = 50                       # requested + used = 50% * available
               = rawScoringFunction(50)
               = 5                        # floor(50/10)

cpu            = resourceScoringFunction((2+1),8)
               = (100 -((8-3)*100/8))
               = 37.5                     # requested + used = 37.5% * available
               = rawScoringFunction(37.5)
               = 3                        # floor(37.5/10)

NodeScore   =  ((7 * 5) + (5 * 1) + (3 * 3)) / (5 + 1 + 3)
            =  5

Node 2 spec:

Available:
  intel.com/foo: 8
  memory: 1GB
  cpu: 8
Used:
  intel.com/foo: 2
  memory: 512MB
  cpu: 6

Node score:

intel.com/foo  = resourceScoringFunction((2+2),8)
               =  (100 - ((8-4)*100/8)
               =  (100 - 50)
               =  50
               =  rawScoringFunction(50)
               = 5

memory         = resourceScoringFunction((256+512),1024)
               = (100 -((1024-768)*100/1024))
               = 75
               = rawScoringFunction(75)
               = 7

cpu            = resourceScoringFunction((2+6),8)
               = (100 -((8-8)*100/8))
               = 100
               = rawScoringFunction(100)
               = 10

NodeScore   =  ((5 * 5) + (7 * 1) + (10 * 3)) / (5 + 1 + 3)
            =  7

What's next

• Read more about the scheduling framework
• Read more about scheduler configuration

11 - Pod Priority and Preemption

Pods can have priority. Priority indicates the importance of a Pod relative to other Pods. If a Pod cannot be scheduled, the scheduler tries to preempt (evict) lower priority Pods to make scheduling of the pending Pod possible.

How to use priority and preemption

To use priority and preemption:

1. Add one or more PriorityClasses.

2. Create Pods withpriorityClassName set to one of the added PriorityClasses. Of course you do not need to create the Pods directly; normally you would add priorityClassName to the Pod template of a collection object like a Deployment.

Keep reading for more information about these steps.

PriorityClass

A PriorityClass is a non-namespaced object that defines a mapping from a priority class name to the integer value of the priority. The name is specified in the name field of the PriorityClass object's metadata. The value is specified in the required value field. The higher the value, the higher the priority. The name of a PriorityClass object must be a valid DNS subdomain name, and it cannot be prefixed with system-.

A PriorityClass object can have any 32-bit integer value smaller than or equal to 1 billion. This means that the range of values for a PriorityClass object is from -2147483648 to 1000000000 inclusive. Larger numbers are reserved for built-in PriorityClasses that represent critical system Pods. A cluster admin should create one PriorityClass object for each such mapping that they want.

PriorityClass also has two optional fields: globalDefault and description. The globalDefault field indicates that the value of this PriorityClass should be used for Pods without a priorityClassName. Only one PriorityClass with globalDefault set to true can exist in the system. If there is no PriorityClass with globalDefault set, the priority of Pods with no priorityClassName is zero.

The description field is an arbitrary string. It is meant to tell users of the cluster when they should use this PriorityClass.

Notes about PodPriority and existing clusters

• If you upgrade an existing cluster without this feature, the priority of your existing Pods is effectively zero.

• Addition of a PriorityClass with globalDefault set to true does not change the priorities of existing Pods. The value of such a PriorityClass is used only for Pods created after the PriorityClass is added.

• If you delete a PriorityClass, existing Pods that use the name of the deleted PriorityClass remain unchanged, but you cannot create more Pods that use the name of the deleted PriorityClass.

Example PriorityClass

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: high-priority
value: 1000000
globalDefault: false
description: "This priority class should be used for XYZ service pods only."

Non-preempting PriorityClass

Pods with preemptionPolicy: Never will be placed in the scheduling queue ahead of lower-priority pods, but they cannot preempt other pods. A non-preempting pod waiting to be scheduled will stay in the scheduling queue, until sufficient resources are free, and it can be scheduled. Non-preempting pods, like other pods, are subject to scheduler back-off. This means that if the scheduler tries these pods and they cannot be scheduled, they will be retried with lower frequency, allowing other pods with lower priority to be scheduled before them.

Non-preempting pods may still be preempted by other, high-priority pods.

preemptionPolicy defaults to PreemptLowerPriority, which will allow pods of that PriorityClass to preempt lower-priority pods (as is existing default behavior). If preemptionPolicy is set to Never, pods in that PriorityClass will be non-preempting.

An example use case is for data science workloads. A user may submit a job that they want to be prioritized above other workloads, but do not wish to discard existing work by preempting running pods. The high priority job with preemptionPolicy: Never will be scheduled ahead of other queued pods, as soon as sufficient cluster resources "naturally" become free.

Example Non-preempting PriorityClass

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: high-priority-nonpreempting
value: 1000000
preemptionPolicy: Never
globalDefault: false
description: "This priority class will not cause other pods to be preempted."

Pod priority

After you have one or more PriorityClasses, you can create Pods that specify one of those PriorityClass names in their specifications. The priority admission controller uses the priorityClassName field and populates the integer value of the priority. If the priority class is not found, the Pod is rejected.

The following YAML is an example of a Pod configuration that uses the PriorityClass created in the preceding example. The priority admission controller checks the specification and resolves the priority of the Pod to 1000000.

apiVersion: v1
kind: Pod
metadata:
  name: nginx
  labels:
    env: test
spec:
  containers:
  - name: nginx
    image: nginx
    imagePullPolicy: IfNotPresent
  priorityClassName: high-priority

Effect of Pod priority on scheduling order

When Pod priority is enabled, the scheduler orders pending Pods by their priority and a pending Pod is placed ahead of other pending Pods with lower priority in the scheduling queue. As a result, the higher priority Pod may be scheduled sooner than Pods with lower priority if its scheduling requirements are met. If such Pod cannot be scheduled, the scheduler will continue and try to schedule other lower priority Pods.

Preemption

When Pods are created, they go to a queue and wait to be scheduled. The scheduler picks a Pod from the queue and tries to schedule it on a Node. If no Node is found that satisfies all the specified requirements of the Pod, preemption logic is triggered for the pending Pod. Let's call the pending Pod P. Preemption logic tries to find a Node where removal of one or more Pods with lower priority than P would enable P to be scheduled on that Node. If such a Node is found, one or more lower priority Pods get evicted from the Node. After the Pods are gone, P can be scheduled on the Node.

User exposed information

When Pod P preempts one or more Pods on Node N, nominatedNodeName field of Pod P's status is set to the name of Node N. This field helps the scheduler track resources reserved for Pod P and also gives users information about preemptions in their clusters.

Please note that Pod P is not necessarily scheduled to the "nominated Node". The scheduler always tries the "nominated Node" before iterating over any other nodes. After victim Pods are preempted, they get their graceful termination period. If another node becomes available while scheduler is waiting for the victim Pods to terminate, scheduler may use the other node to schedule Pod P. As a result nominatedNodeName and nodeName of Pod spec are not always the same. Also, if the scheduler preempts Pods on Node N, but then a higher priority Pod than Pod P arrives, the scheduler may give Node N to the new higher priority Pod. In such a case, scheduler clears nominatedNodeName of Pod P. By doing this, scheduler makes Pod P eligible to preempt Pods on another Node.

Limitations of preemption

Graceful termination of preemption victims

When Pods are preempted, the victims get their graceful termination period. They have that much time to finish their work and exit. If they don't, they are killed. This graceful termination period creates a time gap between the point that the scheduler preempts Pods and the time when the pending Pod (P) can be scheduled on the Node (N). In the meantime, the scheduler keeps scheduling other pending Pods. As victims exit or get terminated, the scheduler tries to schedule Pods in the pending queue. Therefore, there is usually a time gap between the point that scheduler preempts victims and the time that Pod P is scheduled. In order to minimize this gap, one can set graceful termination period of lower priority Pods to zero or a small number.

PodDisruptionBudget is supported, but not guaranteed

A PodDisruptionBudget (PDB) allows application owners to limit the number of Pods of a replicated application that are down simultaneously from voluntary disruptions. Kubernetes supports PDB when preempting Pods, but respecting PDB is best effort. The scheduler tries to find victims whose PDB are not violated by preemption, but if no such victims are found, preemption will still happen, and lower priority Pods will be removed despite their PDBs being violated.

Inter-Pod affinity on lower-priority Pods

A Node is considered for preemption only when the answer to this question is yes: "If all the Pods with lower priority than the pending Pod are removed from the Node, can the pending Pod be scheduled on the Node?"

If a pending Pod has inter-pod affinity to one or more of the lower-priority Pods on the Node, the inter-Pod affinity rule cannot be satisfied in the absence of those lower-priority Pods. In this case, the scheduler does not preempt any Pods on the Node. Instead, it looks for another Node. The scheduler might find a suitable Node or it might not. There is no guarantee that the pending Pod can be scheduled.

Our recommended solution for this problem is to create inter-Pod affinity only towards equal or higher priority Pods.

Cross node preemption

Suppose a Node N is being considered for preemption so that a pending Pod P can be scheduled on N. P might become feasible on N only if a Pod on another Node is preempted. Here's an example:

• Pod P is being considered for Node N.
• Pod Q is running on another Node in the same Zone as Node N.
• Pod P has Zone-wide anti-affinity with Pod Q (topologyKey: topology.kubernetes.io/zone).
• There are no other cases of anti-affinity between Pod P and other Pods in the Zone.
• In order to schedule Pod P on Node N, Pod Q can be preempted, but scheduler does not perform cross-node preemption. So, Pod P will be deemed unschedulable on Node N.

If Pod Q were removed from its Node, the Pod anti-affinity violation would be gone, and Pod P could possibly be scheduled on Node N.

We may consider adding cross Node preemption in future versions if there is enough demand and if we find an algorithm with reasonable performance.

Troubleshooting

Pod priority and preemption can have unwanted side effects. Here are some examples of potential problems and ways to deal with them.

Pods are preempted unnecessarily

Preemption removes existing Pods from a cluster under resource pressure to make room for higher priority pending Pods. If you give high priorities to certain Pods by mistake, these unintentionally high priority Pods may cause preemption in your cluster. Pod priority is specified by setting the priorityClassName field in the Pod's specification. The integer value for priority is then resolved and populated to the priority field of podSpec.

To address the problem, you can change the priorityClassName for those Pods to use lower priority classes, or leave that field empty. An empty priorityClassName is resolved to zero by default.

When a Pod is preempted, there will be events recorded for the preempted Pod. Preemption should happen only when a cluster does not have enough resources for a Pod. In such cases, preemption happens only when the priority of the pending Pod (preemptor) is higher than the victim Pods. Preemption must not happen when there is no pending Pod, or when the pending Pods have equal or lower priority than the victims. If preemption happens in such scenarios, please file an issue.

Pods are preempted, but the preemptor is not scheduled

When pods are preempted, they receive their requested graceful termination period, which is by default 30 seconds. If the victim Pods do not terminate within this period, they are forcibly terminated. Once all the victims go away, the preemptor Pod can be scheduled.

While the preemptor Pod is waiting for the victims to go away, a higher priority Pod may be created that fits on the same Node. In this case, the scheduler will schedule the higher priority Pod instead of the preemptor.

This is expected behavior: the Pod with the higher priority should take the place of a Pod with a lower priority.

Higher priority Pods are preempted before lower priority pods

The scheduler tries to find nodes that can run a pending Pod. If no node is found, the scheduler tries to remove Pods with lower priority from an arbitrary node in order to make room for the pending pod. If a node with low priority Pods is not feasible to run the pending Pod, the scheduler may choose another node with higher priority Pods (compared to the Pods on the other node) for preemption. The victims must still have lower priority than the preemptor Pod.

When there are multiple nodes available for preemption, the scheduler tries to choose the node with a set of Pods with lowest priority. However, if such Pods have PodDisruptionBudget that would be violated if they are preempted then the scheduler may choose another node with higher priority Pods.

When multiple nodes exist for preemption and none of the above scenarios apply, the scheduler chooses a node with the lowest priority.

Interactions between Pod priority and quality of service

Pod priority and QoS class are two orthogonal features with few interactions and no default restrictions on setting the priority of a Pod based on its QoS classes. The scheduler's preemption logic does not consider QoS when choosing preemption targets. Preemption considers Pod priority and attempts to choose a set of targets with the lowest priority. Higher-priority Pods are considered for preemption only if the removal of the lowest priority Pods is not sufficient to allow the scheduler to schedule the preemptor Pod, or if the lowest priority Pods are protected by PodDisruptionBudget.

The kubelet uses Priority to determine pod order for node-pressure eviction. You can use the QoS class to estimate the order in which pods are most likely to get evicted. The kubelet ranks pods for eviction based on the following factors:

1. Whether the starved resource usage exceeds requests
2. Pod Priority
3. Amount of resource usage relative to requests

See Pod selection for kubelet eviction for more details.

kubelet node-pressure eviction does not evict Pods when their usage does not exceed their requests. If a Pod with lower priority is not exceeding its requests, it won't be evicted. Another Pod with higher priority that exceeds its requests may be evicted.

What's next

• Read about using ResourceQuotas in connection with PriorityClasses: limit Priority Class consumption by default
• Learn about Pod Disruption
• Learn about API-initiated Eviction
• Learn about Node-pressure Eviction

12 - Node-pressure Eviction

Node-pressure eviction is the process by which the kubelet proactively terminates pods to reclaim resources on nodes.

The kubelet monitors resources like memory, disk space, and filesystem inodes on your cluster's nodes. When one or more of these resources reach specific consumption levels, the kubelet can proactively fail one or more pods on the node to reclaim resources and prevent starvation.

During a node-pressure eviction, the kubelet sets the phase for the selected pods to Failed, and terminates the Pod.

Node-pressure eviction is not the same as API-initiated eviction.

The kubelet does not respect your configured PodDisruptionBudget or the pod's terminationGracePeriodSeconds. If you use soft eviction thresholds, the kubelet respects your configured eviction-max-pod-grace-period. If you use hard eviction thresholds, the kubelet uses a 0s grace period (immediate shutdown) for termination.

Self healing behavior

The kubelet attempts to reclaim node-level resources before it terminates end-user pods. For example, it removes unused container images when disk resources are starved.

If the pods are managed by a workload management object (such as StatefulSet or Deployment) that replaces failed pods, the control plane (kube-controller-manager) creates new pods in place of the evicted pods.

Self healing for static pods

If you are running a static pod on a node that is under resource pressure, the kubelet may evict that static Pod. The kubelet then tries to create a replacement, because static Pods always represent an intent to run a Pod on that node.

The kubelet takes the priority of the static pod into account when creating a replacement. If the static pod manifest specifies a low priority, and there are higher-priority Pods defined within the cluster's control plane, and the node is under resource pressure, the kubelet may not be able to make room for that static pod. The kubelet continues to attempt to run all static pods even when there is resource pressure on a node.

Eviction signals and thresholds

The kubelet uses various parameters to make eviction decisions, like the following:

• Eviction signals
• Eviction thresholds
• Monitoring intervals

Eviction signals

Eviction signals are the current state of a particular resource at a specific point in time. The kubelet uses eviction signals to make eviction decisions by comparing the signals to eviction thresholds, which are the minimum amount of the resource that should be available on the node.

The kubelet uses the following eviction signals:

In this table, the Description column shows how kubelet gets the value of the signal. Each signal supports either a percentage or a literal value. The kubelet calculates the percentage value relative to the total capacity associated with the signal.

Memory signals

On Linux nodes, the value for memory.available is derived from the cgroupfs instead of tools like free -m. This is important because free -m does not work in a container, and if users use the node allocatable feature, out of resource decisions are made local to the end user Pod part of the cgroup hierarchy as well as the root node. This script or cgroupv2 script reproduces the same set of steps that the kubelet performs to calculate memory.available. The kubelet excludes inactive_file (the number of bytes of file-backed memory on the inactive LRU list) from its calculation, as it assumes that memory is reclaimable under pressure.

On Windows nodes, the value for memory.available is derived from the node's global memory commit levels (queried through the GetPerformanceInfo() system call) by subtracting the node's global CommitTotal from the node's CommitLimit. Please note that CommitLimit can change if the node's page-file size changes!

Filesystem signals

The kubelet recognizes three specific filesystem identifiers that can be used with eviction signals (<identifier>.inodesFree or <identifier>.available):

1. nodefs: The node's main filesystem, used for local disk volumes, emptyDir volumes not backed by memory, log storage, ephemeral storage, and more. For example, nodefs contains /var/lib/kubelet.

2. imagefs: An optional filesystem that container runtimes can use to store container images (which are the read-only layers) and container writable layers.

3. containerfs: An optional filesystem that container runtime can use to store the writeable layers. Similar to the main filesystem (see nodefs), it's used to store local disk volumes, emptyDir volumes not backed by memory, log storage, and ephemeral storage, except for the container images. When containerfs is used, the imagefs filesystem can be split to only store images (read-only layers) and nothing else.

As such, kubelet generally allows three options for container filesystems:

• Everything is on the single nodefs, also referred to as "rootfs" or simply "root", and there is no dedicated image filesystem.

• Container storage (see nodefs) is on a dedicated disk, and imagefs (writable and read-only layers) is separate from the root filesystem. This is often referred to as "split disk" (or "separate disk") filesystem.

• Container filesystem containerfs (same as nodefs plus writable layers) is on root and the container images (read-only layers) are stored on separate imagefs. This is often referred to as "split image" filesystem.

The kubelet will attempt to auto-discover these filesystems with their current configuration directly from the underlying container runtime and will ignore other local node filesystems.

The kubelet does not support other container filesystems or storage configurations, and it does not currently support multiple filesystems for images and containers.

Deprecated kubelet garbage collection features

Some kubelet garbage collection features are deprecated in favor of eviction:

Eviction thresholds

You can specify custom eviction thresholds for the kubelet to use when it makes eviction decisions. You can configure soft and hard eviction thresholds.

Eviction thresholds have the form [eviction-signal][operator][quantity], where:

• eviction-signal is the eviction signal to use.
• operator is the relational operator you want, such as < (less than).
• quantity is the eviction threshold amount, such as 1Gi. The value of quantity must match the quantity representation used by Kubernetes. You can use either literal values or percentages (%).

For example, if a node has 10GiB of total memory and you want trigger eviction if the available memory falls below 1GiB, you can define the eviction threshold as either memory.available<10% or memory.available<1Gi (you cannot use both).

Soft eviction thresholds

A soft eviction threshold pairs an eviction threshold with a required administrator-specified grace period. The kubelet does not evict pods until the grace period is exceeded. The kubelet returns an error on startup if you do not specify a grace period.

You can specify both a soft eviction threshold grace period and a maximum allowed pod termination grace period for kubelet to use during evictions. If you specify a maximum allowed grace period and the soft eviction threshold is met, the kubelet uses the lesser of the two grace periods. If you do not specify a maximum allowed grace period, the kubelet kills evicted pods immediately without graceful termination.

You can use the following flags to configure soft eviction thresholds:

• eviction-soft: A set of eviction thresholds like memory.available<1.5Gi that can trigger pod eviction if held over the specified grace period.
• eviction-soft-grace-period: A set of eviction grace periods like memory.available=1m30s that define how long a soft eviction threshold must hold before triggering a Pod eviction.
• eviction-max-pod-grace-period: The maximum allowed grace period (in seconds) to use when terminating pods in response to a soft eviction threshold being met.

Hard eviction thresholds

A hard eviction threshold has no grace period. When a hard eviction threshold is met, the kubelet kills pods immediately without graceful termination to reclaim the starved resource.

You can use the eviction-hard flag to configure a set of hard eviction thresholds like memory.available<1Gi.

The kubelet has the following default hard eviction thresholds:

• memory.available<100Mi (Linux nodes)
• memory.available<500Mi (Windows nodes)
• nodefs.available<10%
• imagefs.available<15%
• nodefs.inodesFree<5% (Linux nodes)
• imagefs.inodesFree<5% (Linux nodes)

These default values of hard eviction thresholds will only be set if none of the parameters is changed. If you change the value of any parameter, then the values of other parameters will not be inherited as the default values and will be set to zero. In order to provide custom values, you should provide all the thresholds respectively.

The containerfs.available and containerfs.inodesFree (Linux nodes) default eviction thresholds will be set as follows:

• If a single filesystem is used for everything, then containerfs thresholds are set the same as nodefs.

• If separate filesystems are configured for both images and containers, then containerfs thresholds are set the same as imagefs.

Setting custom overrides for thresholds related to containersfs is currently not supported, and a warning will be issued if an attempt to do so is made; any provided custom values will, as such, be ignored.

Eviction monitoring interval

The kubelet evaluates eviction thresholds based on its configured housekeeping-interval, which defaults to 10s.

Node conditions

The kubelet reports node conditions to reflect that the node is under pressure because hard or soft eviction threshold is met, independent of configured grace periods.

The kubelet maps eviction signals to node conditions as follows:

The control plane also maps these node conditions to taints.

The kubelet updates the node conditions based on the configured --node-status-update-frequency, which defaults to 10s.

Node condition oscillation

In some cases, nodes oscillate above and below soft eviction thresholds without holding for the defined grace periods. This causes the reported node condition to constantly switch between true and false, leading to bad eviction decisions.

To protect against oscillation, you can use the eviction-pressure-transition-period flag, which controls how long the kubelet must wait before transitioning a node condition to a different state. The transition period has a default value of 5m.

Reclaiming node level resources

The kubelet tries to reclaim node-level resources before it evicts end-user pods.

When a DiskPressure node condition is reported, the kubelet reclaims node-level resources based on the filesystems on the node.

Without imagefs or containerfs

If the node only has a nodefs filesystem that meets eviction thresholds, the kubelet frees up disk space in the following order:

1. Garbage collect dead pods and containers.
2. Delete unused images.

With imagefs

If the node has a dedicated imagefs filesystem for container runtimes to use, the kubelet does the following:

• If the nodefs filesystem meets the eviction thresholds, the kubelet garbage collects dead pods and containers.

• If the imagefs filesystem meets the eviction thresholds, the kubelet deletes all unused images.

With imagefs and containerfs

If the node has a dedicated containerfs alongside the imagefs filesystem configured for the container runtimes to use, then kubelet will attempt to reclaim resources as follows:

• If the containerfs filesystem meets the eviction thresholds, the kubelet garbage collects dead pods and containers.

• If the imagefs filesystem meets the eviction thresholds, the kubelet deletes all unused images.

Pod selection for kubelet eviction

If the kubelet's attempts to reclaim node-level resources don't bring the eviction signal below the threshold, the kubelet begins to evict end-user pods.

The kubelet uses the following parameters to determine the pod eviction order:

1. Whether the pod's resource usage exceeds requests
2. Pod Priority
3. The pod's resource usage relative to requests

As a result, kubelet ranks and evicts pods in the following order:

1. BestEffort or Burstable pods where the usage exceeds requests. These pods are evicted based on their Priority and then by how much their usage level exceeds the request.

2. Guaranteed pods and Burstable pods where the usage is less than requests are evicted last, based on their Priority.

Guaranteed pods are guaranteed only when requests and limits are specified for all the containers and they are equal. These pods will never be evicted because of another pod's resource consumption. If a system daemon (such as kubelet and journald) is consuming more resources than were reserved via system-reserved or kube-reserved allocations, and the node only has Guaranteed or Burstable pods using less resources than requests left on it, then the kubelet must choose to evict one of these pods to preserve node stability and to limit the impact of resource starvation on other pods. In this case, it will choose to evict pods of lowest Priority first.

If you are running a static pod and want to avoid having it evicted under resource pressure, set the priority field for that Pod directly. Static pods do not support the priorityClassName field.

When the kubelet evicts pods in response to inode or process ID starvation, it uses the Pods' relative priority to determine the eviction order, because inodes and PIDs have no requests.

The kubelet sorts pods differently based on whether the node has a dedicated imagefs or containerfs filesystem:

Without imagefs or containerfs (nodefs and imagefs use the same filesystem)

• If nodefs triggers evictions, the kubelet sorts pods based on their total disk usage (local volumes + logs and a writable layer of all containers).

With imagefs (nodefs and imagefs filesystems are separate)

• If nodefs triggers evictions, the kubelet sorts pods based on nodefs usage (local volumes + logs of all containers).

• If imagefs triggers evictions, the kubelet sorts pods based on the writable layer usage of all containers.

With imagesfs and containerfs (imagefs and containerfs have been split)

• If containerfs triggers evictions, the kubelet sorts pods based on containerfs usage (local volumes + logs and a writable layer of all containers).

• If imagefs triggers evictions, the kubelet sorts pods based on the storage of images rank, which represents the disk usage of a given image.

Minimum eviction reclaim

In some cases, pod eviction only reclaims a small amount of the starved resource. This can lead to the kubelet repeatedly hitting the configured eviction thresholds and triggering multiple evictions.

You can use the --eviction-minimum-reclaim flag or a kubelet config file to configure a minimum reclaim amount for each resource. When the kubelet notices that a resource is starved, it continues to reclaim that resource until it reclaims the quantity you specify.

For example, the following configuration sets minimum reclaim amounts:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
evictionHard:
  memory.available: "500Mi"
  nodefs.available: "1Gi"
  imagefs.available: "100Gi"
evictionMinimumReclaim:
  memory.available: "0Mi"
  nodefs.available: "500Mi"
  imagefs.available: "2Gi"

In this example, if the nodefs.available signal meets the eviction threshold, the kubelet reclaims the resource until the signal reaches the threshold of 1GiB, and then continues to reclaim the minimum amount of 500MiB, until the available nodefs storage value reaches 1.5GiB.

Similarly, the kubelet tries to reclaim the imagefs resource until the imagefs.available value reaches 102Gi, representing 102 GiB of available container image storage. If the amount of storage that the kubelet could reclaim is less than 2GiB, the kubelet doesn't reclaim anything.

The default eviction-minimum-reclaim is 0 for all resources.

Node out of memory behavior

If the node experiences an out of memory (OOM) event prior to the kubelet being able to reclaim memory, the node depends on the oom_killer to respond.

The kubelet sets an oom_score_adj value for each container based on the QoS for the pod.

If the kubelet can't reclaim memory before a node experiences OOM, the oom_killer calculates an oom_score based on the percentage of memory it's using on the node, and then adds the oom_score_adj to get an effective oom_score for each container. It then kills the container with the highest score.

This means that containers in low QoS pods that consume a large amount of memory relative to their scheduling requests are killed first.

Unlike pod eviction, if a container is OOM killed, the kubelet can restart it based on its restartPolicy.

Good practices

The following sections describe good practice for eviction configuration.

Schedulable resources and eviction policies

When you configure the kubelet with an eviction policy, you should make sure that the scheduler will not schedule pods if they will trigger eviction because they immediately induce memory pressure.

Consider the following scenario:

• Node memory capacity: 10GiB
• Operator wants to reserve 10% of memory capacity for system daemons (kernel, kubelet, etc.)
• Operator wants to evict Pods at 95% memory utilization to reduce incidence of system OOM.

For this to work, the kubelet is launched as follows:

--eviction-hard=memory.available<500Mi
--system-reserved=memory=1.5Gi

In this configuration, the --system-reserved flag reserves 1.5GiB of memory for the system, which is 10% of the total memory + the eviction threshold amount.

The node can reach the eviction threshold if a pod is using more than its request, or if the system is using more than 1GiB of memory, which makes the memory.available signal fall below 500MiB and triggers the threshold.

DaemonSets and node-pressure eviction

Pod priority is a major factor in making eviction decisions. If you do not want the kubelet to evict pods that belong to a DaemonSet, give those pods a high enough priority by specifying a suitable priorityClassName in the pod spec. You can also use a lower priority, or the default, to only allow pods from that DaemonSet to run when there are enough resources.

Known issues

The following sections describe known issues related to out of resource handling.

kubelet may not observe memory pressure right away

By default, the kubelet polls cAdvisor to collect memory usage stats at a regular interval. If memory usage increases within that window rapidly, the kubelet may not observe MemoryPressure fast enough, and the OOM killer will still be invoked.

You can use the --kernel-memcg-notification flag to enable the memcg notification API on the kubelet to get notified immediately when a threshold is crossed.

If you are not trying to achieve extreme utilization, but a sensible measure of overcommit, a viable workaround for this issue is to use the --kube-reserved and --system-reserved flags to allocate memory for the system.

active_file memory is not considered as available memory

On Linux, the kernel tracks the number of bytes of file-backed memory on active least recently used (LRU) list as the active_file statistic. The kubelet treats active_file memory areas as not reclaimable. For workloads that make intensive use of block-backed local storage, including ephemeral local storage, kernel-level caches of file and block data means that many recently accessed cache pages are likely to be counted as active_file. If enough of these kernel block buffers are on the active LRU list, the kubelet is liable to observe this as high resource use and taint the node as experiencing memory pressure - triggering pod eviction.

For more details, see https://github.com/kubernetes/kubernetes/issues/43916

You can work around that behavior by setting the memory limit and memory request the same for containers likely to perform intensive I/O activity. You will need to estimate or measure an optimal memory limit value for that container.

What's next

• Learn about API-initiated Eviction
• Learn about Pod Priority and Preemption
• Learn about PodDisruptionBudgets
• Learn about Quality of Service (QoS)
• Check out the Eviction API

13 - API-initiated Eviction

API-initiated eviction is the process by which you use the Eviction API to create an Eviction object that triggers graceful pod termination.

You can request eviction by calling the Eviction API directly, or programmatically using a client of the API server, like the kubectl drain command. This creates an Eviction object, which causes the API server to terminate the Pod.

API-initiated evictions respect your configured PodDisruptionBudgets and terminationGracePeriodSeconds.

Using the API to create an Eviction object for a Pod is like performing a policy-controlled DELETE operation on the Pod.

Calling the Eviction API

You can use a Kubernetes language client to access the Kubernetes API and create an Eviction object. To do this, you POST the attempted operation, similar to the following example:

• policy/v1
• policy/v1beta1

{
  "apiVersion": "policy/v1",
  "kind": "Eviction",
  "metadata": {
    "name": "quux",
    "namespace": "default"
  }
}

{
  "apiVersion": "policy/v1beta1",
  "kind": "Eviction",
  "metadata": {
    "name": "quux",
    "namespace": "default"
  }
}

Alternatively, you can attempt an eviction operation by accessing the API using curl or wget, similar to the following example:

curl -v -H 'Content-type: application/json' https://your-cluster-api-endpoint.example/api/v1/namespaces/default/pods/quux/eviction -d @eviction.json

How API-initiated eviction works

When you request an eviction using the API, the API server performs admission checks and responds in one of the following ways:

• 200 OK: the eviction is allowed, the Eviction subresource is created, and the Pod is deleted, similar to sending a DELETE request to the Pod URL.
• 429 Too Many Requests: the eviction is not currently allowed because of the configured PodDisruptionBudget. You may be able to attempt the eviction again later. You might also see this response because of API rate limiting.
• 500 Internal Server Error: the eviction is not allowed because there is a misconfiguration, like if multiple PodDisruptionBudgets reference the same Pod.

If the Pod you want to evict isn't part of a workload that has a PodDisruptionBudget, the API server always returns 200 OK and allows the eviction.

If the API server allows the eviction, the Pod is deleted as follows:

1. The Pod resource in the API server is updated with a deletion timestamp, after which the API server considers the Pod resource to be terminated. The Pod resource is also marked with the configured grace period.
2. The kubelet on the node where the local Pod is running notices that the Pod resource is marked for termination and starts to gracefully shut down the local Pod.
3. While the kubelet is shutting the Pod down, the control plane removes the Pod from Endpoint and EndpointSlice objects. As a result, controllers no longer consider the Pod as a valid object.
4. After the grace period for the Pod expires, the kubelet forcefully terminates the local Pod.
5. The kubelet tells the API server to remove the Pod resource.
6. The API server deletes the Pod resource.

Troubleshooting stuck evictions

In some cases, your applications may enter a broken state, where the Eviction API will only return 429 or 500 responses until you intervene. This can happen if, for example, a ReplicaSet creates pods for your application but new pods do not enter a Ready state. You may also notice this behavior in cases where the last evicted Pod had a long termination grace period.

If you notice stuck evictions, try one of the following solutions:

• Abort or pause the automated operation causing the issue. Investigate the stuck application before you restart the operation.
• Wait a while, then directly delete the Pod from your cluster control plane instead of using the Eviction API.

What's next

• Learn how to protect your applications with a Pod Disruption Budget.
• Learn about Node-pressure Eviction.
• Learn about Pod Priority and Preemption.