hvpa-controller

README

hvpa-controller

Goals

1. The goal of HVPA is to re-use the upstream components HPA and VPA as much as possible for scaling components horizontally and vertically respectively.
   1. HPA for recommendation and scaling horizontally.
   2. VPA for recommendation for scaling vertically.
2. Where there are gaps in using HPA and VPA simultaneously to scale a given component, introduce functionality to fill those gaps.
   1. HPA and VPA are recommended to be mixed only in case HPA is used for custom/external metrics as mentioned here. But in some scenarios it might make sense to use them both even for CPU and memory (e.g. kube-apiserver or ingress etc.)
   2. VPA updates the pods directly (via webhooks) which may confuse HPA (especially, when used for the same scaling metrics) as it may not see the changes in the upstream targetRefs.
3. Where there is functionality missing in either HPA or VPA, introduce them to provide more flexibility during horizontal and vertical scaling. Especially, for the components that experience disruption during scaling.
   1. Weight-based scaling horizontally and vertically silmultaneously.
   2. Support for configurable (at the resource level) threshold levels to trigger VPA (and possibly HPA) updates to minimize unnecessary scaling of components. Especially, if scaling is disruptive.
   3. Support for configurable (at the resource level) stabilisation window in all the four directions (up/down/out/in) to stabilize scaling of components. Especially, if scaling is disruptive.
   4. Support for configurable maintenance window for scaling (especially, scaling in/down) for components that do not scale well smoothly (mainly, etcd, but to a lesser extent kube-apiserver as well for WATCH requests). This could be as an alternative or complementary to the stabilisation window mentioned above.
   5. Support for flexible update policy for all four scaling directions (Off/Auto/ScaleUp). ScaleUp would only apply scale up and not scale down (vertically or horizontally). This is again from the perspective of components which experience disruption while scaling (mainly, etcd, but to a lesser extent kube-apiserver as well for WATCH requests). For such components, a ScaleUp update policy will ensure that the component can scale up (with some disruption) automatically to meet the workload requirement but not scale down to avoid unnecessary disruption. This would mean over-provisioning for workloads that experience a short upsurge.
   6. Custom resource support for VPA updates.
      • At the beginning, this would be with some well-defined annotations to describe which part of the CRD to update the resource requests.
      • But eventually, we can think of proposing a Resource subresource (per container) along the lines of Scale subresource.

Non-goals

• It is not a goal of HVPA to duplicate the functionality of the upstream components like HPA and VPA.

Weight based scaling

In this mode, deployment's scaling is divided into stages depending on number of replicas. A user can provide an array of intervals x1 to x2 and corresponding vpaWeights between and including 0 and 1. x1 and x2 are the number of replicas where that stage begins and ends respectively. In a particular stage, HVPA controller will consider VPA's recommendations according to the weights provided, and scale the deployment accordingly. The weights given to the HPA's recommendation will be 1 - vpaWeight.

According to VPA, new requests will be currentRequest + (targetRequest - currentRequest) * vpaWeight

According to HPA, new number of replicas will be:

• ceil of currentReplicas + (desiredReplicas - currentReplicas) * (1 - vpaWeight) if desired replicas is more than current replicas
• floor of currentReplicas + (desiredReplicas - currentReplicas) * (1 - vpaWeight) if desired replicas is less than current replicas

vpaWeight = 1 will result in pure vertical scaling vpaWeight = 0 will result in pure horizontal scaling

Resource request vs number of replicas curve could typically look like:
Here First stage Vpa weight = 0, second stage Vpa weight is a fraction, third stage Vpa weight = 0

resource
request
^
|
|                  ____________
|                 /|
|                /
|               /  |
|              /
|  ___________/    |
|
|-------------|----|-----------|----->
   min        x1    x2        max     #Replicas

Example of vpaWeight vs number of replicas.
Here first stage Vpa weight = 0, second stage Vpa weight = 0.4, third stage Vpa weight = 0

vpaWeight
   ^
   |
0.4|           |-------|
   |           |       |
  0|   ________|       |____________
   |------------------------------------->
               x1     x2               #Replicas

Scaling of limits:

HVPA will scale limits also along with requests based on following criteria:

1. If originalLimit is not set, don't set limit.
2. If originalLimit is set but originalRequest not set - K8s will treat the pod as if they were equal, then set limit equal to request
3. If originalLimit and originalRequest are set and if they are equal, recommend limit equal to request.
4. If limit scaling parameters are not set in hvpa.spec.vpa then scale the limits proportionaly as done by VPA
5. If the scaling parameters are provided, then choose the max of the 2 possible values in hvpa.spec.vpa.limitsRequestsGapScaleParams (percentage and value)

Pros

• Works even if HPA and VPA act on different metrices
• Gives better control to user on scaling of apps

Cons

• Need to define behavior when:
  • current number of replicas is x2, the deployment has not yet scaled up to maxAllowed, and user has chosen to scale only horizontally after x2. In this case, the deployment cannot scale up vertically anymore if the load increases.
  • current number of replicas is x1, the deployment has not yet scaled down to minAllowed, and user has chosen to scale only horizontally for lower values than x1. In this case, the deployment cannot scale down vertically anymore if the load decreases.
• Will need both VPA and HPA to be deployed in a recommendation-only mode.

Mitigation

• x1 and x2 will be optional, If either or both of them are not provided, vertical scaling will be done until minAllowed and/or maxAllowed correspondingly.
• Once replica count reaches maxReplicas, and VPA still recommends higher resource requirements, then vertical scaling will be done unconditionally

Spec

Here is the spec for HVPA controller

Example 1:

apiVersion: autoscaling.k8s.io/v1alpha1
kind: Hvpa
metadata:
  name: hvpa-sample
spec:
  maintenanceTimeWindow:
    begin: "220000-0000"
    end:  "230000-0000"
  hpa:
    deploy: true
    selector:
      matchLabels:
        key1: value1
    scaleUp:
      updatePolicy:
        updateMode: "Auto"
      stabilizationDuration: "2m"
    scaleDown:
      updatePolicy:
        updateMode: "Auto"
      stabilizationDuration: "3m"
    template:
      metadata:
        labels:
          key1: value1
      spec:
        maxReplicas: 10
        minReplicas: 1
        metrics:
        - resource:
            name: memory
            targetAverageUtilization: 50
          type: Resource
        - resource:
            name: cpu
            targetAverageUtilization: 50
          type: Resource
 vpa:
    deploy: true
    selector:
      matchLabels:
        key2: value2
    scaleUp:
      updatePolicy:
        updateMode: "Auto"
      stabilizationDuration: "2m"
      minChange:
        cpu:
          value: "3"
          percentage: 80
        memory:
          value: "3"
          percentage: 80
    scaleDown:
      updatePolicy:
        updateMode: "MaintenanceWindow"
      stabilizationDuration: "3m"
      minChange:
        cpu:
          value: "3"
          percentage: 80
        memory:
          value: "3"
          percentage: 80
    limitsRequestsGapScaleParams:
      cpu:
        percentage: 30
        value: "1"
      memory:
        percentage: 30
        value: 1G
    template:
      metadata:
        labels:
          key2: value2
      spec:
        resourcePolicy:
          containerPolicies:
            - containerName: resource-consumer
              minAllowed:
                memory: 400Mi
              maxAllowed:
                memory: 3000Mi
  weightBasedScalingIntervals:
    - vpaWeight: 0
      startReplicaCount: 1
      lastReplicaCount: 2
    - vpaWeight: 0.6
      startReplicaCount: 3
      lastReplicaCount: 7
    - vpaWeight: 0
      startReplicaCount: 7
      lastReplicaCount: 10
  targetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: resource-consumer

resource
request
^
|
|                  _________
|                 /|
|                /
|               /  |
|              /
|  ___________/    |
|
|-------------|----|--------|----->
   1          3    7        10     #Replicas

Example 2:

apiVersion: autoscaling.k8s.io/v1alpha1
kind: Hvpa
metadata:
  name: hvpa-sample
spec:
  weightBasedScalingIntervals:
    - vpaWeight: 1
      startReplicaCount: 1
      lastReplicaCount: 1
    - vpaWeight: 0.6
      startReplicaCount: 2
      lastReplicaCount: 4
    - vpaWeight: 0
      startReplicaCount: 5
      lastReplicaCount: 10
  hpa:
    .
    .
    .

resource
request
^
|       ______
|      /
|     /  
|    /
|   /  
|  |
|  |  
|  |
|
|--|----|-----|------->
   1   4       10     #Replicas

Example 3: After maxReplicas, even if weight is not mentioned, there is only vertical scaling because HPA is limited by maxReplicas

apiVersion: autoscaling.k8s.io/v1alpha1
kind: Hvpa
metadata:
  name: hvpa-sample
spec:
  weightBasedScalingIntervals:
    - vpaWeight: 0
      startReplicaCount: 1
      lastReplicaCount: 3
    - vpaWeight: 0.6
      startReplicaCount: 4
      lastReplicaCount: 10
  hpa:
    .
    .
    template:
      .
      .
      spec:
        minReplicas: 1
        maxReplicas: 10
        metrics:
        .
        .
    .
    .
    .

resource
request
^
|
|           |
|           |
|           .     
|          /
|         /
|        / 
|       /
|  ____/ 
|
|--|---|----|------>
   1   3    10     #Replicas

Custom Resource Definition

Types

See here.

Example

See here.

Development

Configure envtest binaries to 1.24.2 before you run make test:

source <(setup-envtest use -p env 1.24.2)