k8s.io/autoscaler/vertical-pod-autoscaler

Vertical Pod Autoscaler

Contents

• Contents
• Intro
• Installation
  • Compatibility
  • Notice on deprecation of v1beta2 version (>=0.13.0)
  • Notice on removal of v1beta1 version (>=0.5.0)
  • Prerequisites
  • Install command
  • Quick start
  • Test your installation
  • Example VPA configuration
  • Troubleshooting
  • Components of VPA
  • Tear down
• Limits control
• Examples
  • Keeping limit proportional to request
  • Capping to Limit Range
  • Resource Policy Overriding Limit Range
  • Starting multiple recommenders
  • Using CPU management with static policy
  • Controlling eviction behavior based on scaling direction and resource
  • Limiting which namespaces are used
• Known limitations
• Related links

Intro

Vertical Pod Autoscaler (VPA) frees users from the necessity of setting up-to-date resource limits and requests for the containers in their pods. When configured, it will set the requests automatically based on usage and thus allow proper scheduling onto nodes so that appropriate resource amount is available for each pod. It will also maintain ratios between limits and requests that were specified in initial containers configuration.

It can both down-scale pods that are over-requesting resources, and also up-scale pods that are under-requesting resources based on their usage over time.

Autoscaling is configured with a Custom Resource Definition object called VerticalPodAutoscaler. It allows to specify which pods should be vertically autoscaled as well as if/how the resource recommendations are applied.

To enable vertical pod autoscaling on your cluster please follow the installation procedure described below.

Installation

The current default version is Vertical Pod Autoscaler 1.2.1

Compatibility

VPA version	Kubernetes version
1.2.1	1.27+
1.2.0	1.27+
1.1.2	1.25+
1.1.1	1.25+
1.0	1.25+
0.14	1.25+
0.13	1.25+
0.12	1.25+
0.11	1.22 - 1.24
0.10	1.22+
0.9	1.16+
0.8	1.13+
0.4 to 0.7	1.11+
0.3.X and lower	1.7+

Notice on CRD update (>=1.0.0)

NOTE: In version 1.0.0, we have updated the CRD definition and added RBAC for the status resource. If you are upgrading from version (<=0.14.0), you must update the CRD definition and RBAC.

kubectl apply -f https://raw.githubusercontent.com/kubernetes/autoscaler/vpa-release-1.0/vertical-pod-autoscaler/deploy/vpa-v1-crd-gen.yaml
kubectl apply -f https://raw.githubusercontent.com/kubernetes/autoscaler/vpa-release-1.0/vertical-pod-autoscaler/deploy/vpa-rbac.yaml

Another method is to re-execute the ./hack/vpa-process-yamls.sh script.

git clone https://github.com/kubernetes/autoscaler.git
cd autoscaler/vertical-pod-autoscaler
git checkout origin/vpa-release-1.0
REGISTRY=registry.k8s.io/autoscaling TAG=1.0.0 ./hack/vpa-process-yamls.sh apply

If you need to roll back to version (<=0.14.0), please check out the release for your rollback version and execute ./hack/vpa-process-yamls.sh. For example, to rollback to 0.14.0:

git checkout origin/vpa-release-0.14
REGISTRY=registry.k8s.io/autoscaling TAG=0.14.0 ./hack/vpa-process-yamls.sh apply
kubectl delete clusterrole system:vpa-status-actor
kubectl delete clusterrolebinding system:vpa-status-actor

Notice on deprecation of v1beta2 version (>=0.13.0)

NOTE: In 0.13.0 we deprecate autoscaling.k8s.io/v1beta2 API. We plan to remove this API version. While for now you can continue to use v1beta2 API we recommend using autoscaling.k8s.io/v1 instead. v1 and v1beta2 APIs are almost identical (v1 API has some fields which are not present in v1beta2) so simply changing which API version you're calling should be enough in almost all cases.

Notice on removal of v1beta1 version (>=0.5.0)

NOTE: In 0.5.0 we disabled the old version of the API - autoscaling.k8s.io/v1beta1. The VPA objects in this version will no longer receive recommendations and existing recommendations will be removed. The objects will remain present though and a ConfigUnsupported condition will be set on them.

This doc is for installing latest VPA. For instructions on migration from older versions see Migration Doc.

Prerequisites

• kubectl should be connected to the cluster you want to install VPA.
• The metrics server must be deployed in your cluster. Read more about Metrics Server.
• If you are using a GKE Kubernetes cluster, you will need to grant your current Google identity cluster-admin role. Otherwise, you won't be authorized to grant extra privileges to the VPA system components.

  $ gcloud info | grep Account    # get current google identity
  Account: [[e-mail]]
  
  $ kubectl create clusterrolebinding myname-cluster-admin-binding --clusterrole=cluster-admin --user=[e-mail]
  Clusterrolebinding "myname-cluster-admin-binding" created
  

• If you already have another version of VPA installed in your cluster, you have to tear down the existing installation first with:

  ./hack/vpa-down.sh
  

Install command

To install VPA, please download the source code of VPA (for example with git clone https://github.com/kubernetes/autoscaler.git) and run the following command inside the vertical-pod-autoscaler directory:

./hack/vpa-up.sh

Note: the script currently reads environment variables: $REGISTRY and $TAG. Make sure you leave them unset unless you want to use a non-default version of VPA.

Note: If you are seeing following error during this step:

unknown option -addext

please upgrade openssl to version 1.1.1 or higher (needs to support -addext option) or use ./hack/vpa-up.sh on the 0.8 release branch.

The script issues multiple kubectl commands to the cluster that insert the configuration and start all needed pods (see architecture) in the kube-system namespace. It also generates and uploads a secret (a CA cert) used by VPA Admission Controller when communicating with the API server.

To print YAML contents with all resources that would be understood by kubectl diff|apply|... commands, you can use

./hack/vpa-process-yamls.sh print

The output of that command won't include secret information generated by pkg/admission-controller/gencerts.sh script.

Quick start

After installation the system is ready to recommend and set resource requests for your pods. In order to use it, you need to insert a Vertical Pod Autoscaler resource for each controller that you want to have automatically computed resource requirements. This will be most commonly a Deployment. There are four modes in which VPAs operate:

• "Auto": VPA assigns resource requests on pod creation as well as updates them on existing pods using the preferred update mechanism. Currently, this is equivalent to "Recreate" (see below). Once restart free ("in-place") update of pod requests is available, it may be used as the preferred update mechanism by the "Auto" mode.
• "Recreate": VPA assigns resource requests on pod creation as well as updates them on existing pods by evicting them when the requested resources differ significantly from the new recommendation (respecting the Pod Disruption Budget, if defined). This mode should be used rarely, only if you need to ensure that the pods are restarted whenever the resource request changes. Otherwise, prefer the "Auto" mode which may take advantage of restart-free updates once they are available.
• "Initial": VPA only assigns resource requests on pod creation and never changes them later.
• "Off": VPA does not automatically change the resource requirements of the pods. The recommendations are calculated and can be inspected in the VPA object.

Test your installation

A simple way to check if Vertical Pod Autoscaler is fully operational in your cluster is to create a sample deployment and a corresponding VPA config:

kubectl create -f examples/hamster.yaml

The above command creates a deployment with two pods, each running a single container that requests 100 millicores and tries to utilize slightly above 500 millicores. The command also creates a VPA config pointing at the deployment. VPA will observe the behaviour of the pods, and after about 5 minutes, they should get updated with a higher CPU request (note that VPA does not modify the template in the deployment, but the actual requests of the pods are updated). To see VPA config and current recommended resource requests run:

kubectl describe vpa

Note: if your cluster has little free capacity these pods may be unable to schedule. You may need to add more nodes or adjust examples/hamster.yaml to use less CPU.

Example VPA configuration

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: my-app-vpa
spec:
  targetRef:
    apiVersion: "apps/v1"
    kind:       Deployment
    name:       my-app
  updatePolicy:
    updateMode: "Auto"

Troubleshooting

To diagnose problems with a VPA installation, perform the following steps:

• Check if all system components are running:

kubectl --namespace=kube-system get pods|grep vpa

The above command should list 3 pods (recommender, updater and admission-controller) all in state Running.

• Check if the system components log any errors. For each of the pods returned by the previous command do:

kubectl --namespace=kube-system logs [pod name] | grep -e '^E[0-9]\{4\}'

• Check that the VPA Custom Resource Definition was created:

kubectl get customresourcedefinition | grep verticalpodautoscalers

Components of VPA

The project consists of 3 components:

• Recommender - it monitors the current and past resource consumption and, based on it, provides recommended values for the containers' cpu and memory requests.

• Updater - it checks which of the managed pods have correct resources set and, if not, kills them so that they can be recreated by their controllers with the updated requests.

• Admission Plugin - it sets the correct resource requests on new pods (either just created or recreated by their controller due to Updater's activity).

More on the architecture can be found HERE.

Tear down

Note that if you stop running VPA in your cluster, the resource requests for the pods already modified by VPA will not change, but any new pods will get resources as defined in your controllers (i.e. deployment or replicaset) and not according to previous recommendations made by VPA.

To stop using Vertical Pod Autoscaling in your cluster:

• If running on GKE, clean up role bindings created in Prerequisites:

kubectl delete clusterrolebinding myname-cluster-admin-binding

• Tear down VPA components:

./hack/vpa-down.sh

Limits control

When setting limits VPA will conform to resource policies. It will maintain limit to request ratio specified for all containers.

VPA will try to cap recommendations between min and max of limit ranges. If limit range conflicts with VPA resource policy, VPA will follow VPA policy (and set values outside the limit range).

To disable getting VPA recommendations for an individual container, set mode to "Off" in containerPolicies.

Examples

Keeping limit proportional to request

The container template specifies resource request for 500 milli CPU and 1 GB of RAM. The template also specifies resource limit of 2 GB RAM. VPA recommendation is 1000 milli CPU and 2 GB of RAM. When VPA applies the recommendation, it will also set the memory limit to 4 GB.

Capping to Limit Range

The container template specifies resource request for 500 milli CPU and 1 GB of RAM. The template also specifies resource limit of 2 GB RAM. A limit range sets a maximum limit to 3 GB RAM per container. VPA recommendation is 1000 milli CPU and 2 GB of RAM. When VPA applies the recommendation, it will set the memory limit to 3 GB (to keep it within the allowed limit range) and the memory request to 1.5 GB ( to maintain a 2:1 limit/request ratio from the template).

Resource Policy Overriding Limit Range

The container template specifies resource request for 500 milli CPU and 1 GB of RAM. The template also specifies a resource limit of 2 GB RAM. A limit range sets a maximum limit to 3 GB RAM per container. VPAs Container Resource Policy requires VPA to set containers request to at least 750 milli CPU and 2 GB RAM. VPA recommendation is 1000 milli CPU and 2 GB of RAM. When applying the recommendation, VPA will set RAM request to 2 GB (following the resource policy) and RAM limit to 4 GB (to maintain the 2:1 limit/request ratio from the template).

Starting multiple recommenders

It is possible to start one or more extra recommenders in order to use different percentile on different workload profiles. For example you could have 3 profiles: frugal, standard and performance which will use different TargetCPUPercentile (50, 90 and 95) to calculate their recommendations.

Please note the usage of the following arguments to override default names and percentiles:

• --recommender-name=performance
• --target-cpu-percentile=0.95

You can then choose which recommender to use by setting recommenders inside the VerticalPodAutoscaler spec.

Custom memory bump-up after OOMKill

After an OOMKill event was observed, VPA increases the memory recommendation based on the observed memory usage in the event according to this formula: recommendation = memory-usage-in-oomkill-event + max(oom-min-bump-up-bytes, memory-usage-in-oomkill-event * oom-bump-up-ratio). You can configure the minimum bump-up as well as the multiplier by specifying startup arguments for the recommender: oom-bump-up-ratio specifies the memory bump up ratio when OOM occurred, default is 1.2. This means, memory will be increased by 20% after an OOMKill event. oom-min-bump-up-bytes specifies minimal increase of memory after observing OOM. Defaults to 100 * 1024 * 1024 (=100MiB)

Usage in recommender deployment

  containers:
  - name: recommender
    args:
      - --oom-bump-up-ratio=2.0
      - --oom-min-bump-up-bytes=524288000

Using CPU management with static policy

If you are using the CPU management with static policy for some containers, you probably want the CPU recommendation to be an integer. A dedicated recommendation pre-processor can perform a round up on the CPU recommendation. Recommendation capping still applies after the round up. To activate this feature, pass the flag --cpu-integer-post-processor-enabled when you start the recommender. The pre-processor only acts on containers having a specific configuration. This configuration consists in an annotation on your VPA object for each impacted container. The annotation format is the following:

vpa-post-processor.kubernetes.io/{containerName}_integerCPU=true

Controlling eviction behavior based on scaling direction and resource

To limit disruptions caused by evictions, you can put additional constraints on the Updater's eviction behavior by specifying .updatePolicy.EvictionRequirements in the VPA spec. An EvictionRequirement contains a resource and a ChangeRequirement, which is evaluated by comparing a new recommendation against the currently set resources for a container. Here is an example configuration which allows evictions only when CPU or memory get scaled up, but not when they both are scaled down

updatePolicy:
  evictionRequirements:
    - resources: ["cpu", "memory"]
      changeRequirement: TargetHigherThanRequests

Note that this doesn't prevent scaling down entirely, as Pods may get recreated for different reasons, resulting in a new recommendation being applied. See the original AEP for more context and usage information.

Limiting which namespaces are used

By default the VPA will run against all namespaces. You can limit that behaviour by setting the following options:

1. ignored-vpa-object-namespaces - A comma separated list of namespaces to ignore
2. vpa-object-namespace - A single namespace to monitor

These options cannot be used together and are mutually exclusive.

Known limitations

• Whenever VPA updates the pod resources, the pod is recreated, which causes all running containers to be recreated. The pod may be recreated on a different node.
• VPA cannot guarantee that pods it evicts or deletes to apply recommendations (when configured in Auto and Recreate modes) will be successfully recreated. This can be partly addressed by using VPA together with Cluster Autoscaler.
• VPA does not update resources of pods which are not run under a controller.
• Vertical Pod Autoscaler should not be used with the Horizontal Pod Autoscaler (HPA) on the same resource metric (CPU or memory) at this moment. However, you can use VPA with HPA on separate resource metrics (e.g. VPA on memory and HPA on CPU) as well as with HPA on custom and external metrics.
• The VPA admission controller is an admission webhook. If you add other admission webhooks to your cluster, it is important to analyze how they interact and whether they may conflict with each other. The order of admission controllers is defined by a flag on API server.
• VPA reacts to most out-of-memory events, but not in all situations.
• VPA performance has not been tested in large clusters.
• VPA recommendation might exceed available resources (e.g. Node size, available size, available quota) and cause pods to go pending. This can be partly addressed by using VPA together with Cluster Autoscaler.
• Multiple VPA resources matching the same pod have undefined behavior.

Related links

• FAQ
• Design proposal
• API definition