关于kubernetes:HPA通过behavior自定义伸缩策略

37次阅读

共计 11716 个字符，预计需要花费 30 分钟才能阅读完成。

k8s v1.18 减少了 hpa v2beta2 的 behavior 字段，能够更精细化的管制伸缩的行为：

  • 若不指定 behavior 字段，则按默认的 behavior 行为执行伸缩；
  • 若指定 behavior 字段，则按自定义的 behavior 行为执行伸缩；

一. demo

若 behavior 的策略 (包含冷却工夫 + 伸缩策略) 不满足需要，能够通过自定义 behavior 精细化管制伸缩的策略。

比方上面的 behavior:

apiVersion: autoscaling/v2beta2
kind: HorizontalPodAutoscaler
metadata:
  name: sample-app
spec:
  ...
  behavior:
    scaleUp:
      policies:
      - type: Percent
        value: 900
        periodSeconds: 300
    scaleDown:
      stabilizationWindowSeconds: 60
      policies:
      - type: Pods
        value: 1
        periodSeconds: 10

  • 扩容时：

    • 立刻扩容；
    • 每次扩容最大(1+9)*currentReplicas，即 10 倍的 replicas；
    • 1 次扩容后，冷却 300s 后能力持续扩容；

  • 缩容时：

    • 冷却 60s 才进行缩容，每次缩容 1 个正本；
    • 1 次缩容后，冷却 10s 后能力持续缩容；

以下面的 Hpa v2beta2 的定义为例，查看其扩缩容的过程:

扩容，将指标猛增(1–>13)：

  • 首先，依照指标计算，应该将正本数从 1 扩容到 13；但因为 scaleUp.policies 的限度，最多扩容 10 倍，即 1 –>10 个正本；
  • 而后，依据指标计算，冷却 300s 后，最终将正本数扩容至 13；
# kubectl describe hpa
Name:                    sample-app
Namespace:               default
Labels:                  <none>
Annotations:             <none>
Reference:               Deployment/sample-app
Metrics:                 (current / target)
  "metric_hpa" on pods:  1 / 1
Min replicas:            1
Max replicas:            15
Behavior:
  Scale Up:
    Stabilization Window: 0 seconds
    Select Policy: Max
    Policies:
      - Type: Percent  Value: 900  Period: 300 seconds
  Scale Down:
    Stabilization Window: 60 seconds
    Select Policy: Max
    Policies:
      - Type: Pods  Value: 1  Period: 10 seconds
Deployment pods:    13 current / 13 desired
Conditions:
  Type            Status  Reason              Message
  ----            ------  ------              -------
  AbleToScale     True    ReadyForNewScale    recommended size matches current size
  ScalingActive   True    ValidMetricFound    the HPA was able to successfully calculate a replica count from pods metric metric_hpa
  ScalingLimited  False   DesiredWithinRange  the desired count is within the acceptable range
Events:
  Type    Reason             Age    From                       Message
  ----    ------             ----   ----                       -------
  Normal  SuccessfulRescale  6m29s  horizontal-pod-autoscaler  New size: 10; reason: pods metric metric_hpa above target
  Normal  SuccessfulRescale  79s    horizontal-pod-autoscaler  New size: 13; reason: pods metric metric_hpa above target

缩容，将指标猛降(13–>1):

  • 首先，冷却 60s 后进行缩容，每次缩容 1 个正本；
  • 而后，待上次缩容后 10s，再次缩容 1 个正本；
  • 最终，缩容至 1 个正本；
# kubectl describe hpa
Name:                    sample-app
Namespace:               default
Labels:                  <none>
Annotations:             <none>
Reference:               Deployment/sample-app
Metrics:                 (current / target)
  "metric_hpa" on pods:  1 / 1
Min replicas:            1
Max replicas:            15
Behavior:
  Scale Up:
    Stabilization Window: 0 seconds
    Select Policy: Max
    Policies:
      - Type: Percent  Value: 900  Period: 300 seconds
  Scale Down:
    Stabilization Window: 60 seconds
    Select Policy: Max
    Policies:
      - Type: Pods  Value: 1  Period: 10 seconds
Deployment pods:    1 current / 1 desired
Conditions:
  Type            Status  Reason              Message
  ----            ------  ------              -------
  AbleToScale     True    ReadyForNewScale    recommended size matches current size
  ScalingActive   True    ValidMetricFound    the HPA was able to successfully calculate a replica count from pods metric metric_hpa
  ScalingLimited  False   DesiredWithinRange  the desired count is within the acceptable range
Events:
  Type    Reason             Age                From                       Message
  ----    ------             ----               ----                       -------
  Normal  SuccessfulRescale  12m                horizontal-pod-autoscaler  New size: 10; reason: pods metric metric_hpa above target
  Normal  SuccessfulRescale  6m59s              horizontal-pod-autoscaler  New size: 13; reason: pods metric metric_hpa above target
  Normal  SuccessfulRescale  3m25s              horizontal-pod-autoscaler  New size: 12; reason: All metrics below target
  Normal  SuccessfulRescale  3m9s               horizontal-pod-autoscaler  New size: 11; reason: All metrics below target
  Normal  SuccessfulRescale  2m54s              horizontal-pod-autoscaler  New size: 10; reason: All metrics below target
  Normal  SuccessfulRescale  2m38s              horizontal-pod-autoscaler  New size: 9; reason: All metrics below target
  Normal  SuccessfulRescale  2m23s              horizontal-pod-autoscaler  New size: 8; reason: All metrics below target
  Normal  SuccessfulRescale  2m7s               horizontal-pod-autoscaler  New size: 7; reason: All metrics below target
  Normal  SuccessfulRescale  112s               horizontal-pod-autoscaler  New size: 6; reason: All metrics below target
  Normal  SuccessfulRescale  34s (x5 over 96s)  horizontal-pod-autoscaler  (combined from similar events): New size: 1; reason: All metrics below target

二. 源码剖析

整个过程分以下几步：

  • 首先，若未设置 scaleDown 的冷却工夫，则配置默认冷却工夫 =300s；
  • 而后，依据 scaleUp/scaleDown 的 stabilizationWindow，计算指标正本数；
  • 最初，依据 scaleUp/scaleDown 的 Policies，计算指标正本数；
// pkg/controller/podautoscaler/horizontal.go
func (a *HorizontalController) normalizeDesiredReplicasWithBehaviors(hpa *autoscalingv2.HorizontalPodAutoscaler, key string, currentReplicas, prenormalizedDesiredReplicas, minReplicas int32) int32 {
    // 1. 设置 scaleDown 的默认冷却工夫
    a.maybeInitScaleDownStabilizationWindow(hpa)
    normalizationArg := NormalizationArg{
        Key:               key,
        ScaleUpBehavior:   hpa.Spec.Behavior.ScaleUp,
        ScaleDownBehavior: hpa.Spec.Behavior.ScaleDown,
        MinReplicas:       minReplicas,
        MaxReplicas:       hpa.Spec.MaxReplicas,
        CurrentReplicas:   currentReplicas,
        DesiredReplicas:   prenormalizedDesiredReplicas}
    // 2. 依据冷却工夫计算正本数
    stabilizedRecommendation, reason, message := a.stabilizeRecommendationWithBehaviors(normalizationArg)
    normalizationArg.DesiredReplicas = stabilizedRecommendation
    ...
    // 3. 依据策略计算正本数
    desiredReplicas, reason, message := a.convertDesiredReplicasWithBehaviorRate(normalizationArg)
    ...
    return desiredReplicas
}

1. 设置 scaleDown 的默认冷却工夫

若 scaleDown.StabilizationWindowSeoncds 未设置，则默认 =300s；

// pkg/controller/podautoscaler/horizontal.go
func (a *HorizontalController) maybeInitScaleDownStabilizationWindow(hpa *autoscalingv2.HorizontalPodAutoscaler) {
    behavior := hpa.Spec.Behavior
    if behavior != nil && behavior.ScaleDown != nil && behavior.ScaleDown.StabilizationWindowSeconds == nil {stabilizationWindowSeconds := (int32)(a.downscaleStabilisationWindow.Seconds())        // 默认 =300s
        hpa.Spec.Behavior.ScaleDown.StabilizationWindowSeconds = &stabilizationWindowSeconds
    }
}

2. 依据冷却工夫计算正本数

依据 stabilizationWindow 计算指标正本数，最终返回 recommendation:

  • 首先，初始值 = 上一步计算的正本数(即指标计算的正本数)；

  • 而后：

    • 若扩容，则 recommendation=min(最近 stabilizationWindowSeconds 的伸缩正本数)，这也意味着冷却 stabilizationWindowSeconds；
    • 若缩容，则 recommendation=max(最近 stabilizationWindowSeconds 的伸缩正本数)，这也意味着冷却 stabilizationWindowSeconds；
// pkg/controller/podautoscaler/horizontal.go
func (a *HorizontalController) stabilizeRecommendationWithBehaviors(args NormalizationArg) (int32, string, string) {
    recommendation := args.DesiredReplicas
    ...
    var betterRecommendation func(int32, int32) int32
    // 扩容
    if args.DesiredReplicas >= args.CurrentReplicas {
        scaleDelaySeconds = *args.ScaleUpBehavior.StabilizationWindowSeconds
       betterRecommendation = min    // min 函数
        reason = "ScaleUpStabilized"
        message = "recent recommendations were lower than current one, applying the lowest recent recommendation"
    } else {    // 缩容
        scaleDelaySeconds = *args.ScaleDownBehavior.StabilizationWindowSeconds
       betterRecommendation = max    // max 函数
        reason = "ScaleDownStabilized"
        message = "recent recommendations were higher than current one, applying the highest recent recommendation"
    }
    ...
    cutoff := time.Now().Add(-time.Second * time.Duration(scaleDelaySeconds))
    for i, rec := range a.recommendations[args.Key] {if rec.timestamp.After(cutoff) {recommendation = betterRecommendation(rec.recommendation, recommendation)
        }
        ...
    }
    ...
    return recommendation, reason, message
}

3. 依据 policies 计算正本数

依据 policies 计算指标正本数

  • 若是扩容：

    • 依据 scaleUp.Policies(percent/pods/period)计算扩容下限；
    • 扩容下限必须 <= hpaMaxReplicas；
    • 最终扩容正本数必须 <= 上一步计算的 desiredReplicas；

  • 若是缩容：

    • 依据 scaleDown.Policies(percent/pods/period)计算缩容下限；
    • 缩容下限必须 >= hpaMinReplicas；
    • 最终缩容正本数必须 >= 上一步计算的 desiredReplicas；
// pkg/controller/podautoscaler/horizontal.go
func (a *HorizontalController) convertDesiredReplicasWithBehaviorRate(args NormalizationArg) (int32, string, string) {
    var possibleLimitingReason, possibleLimitingMessage string
    // 扩容
    if args.DesiredReplicas > args.CurrentReplicas {
        // 依据 scaleUp.Policies 计算扩容下限
        scaleUpLimit := calculateScaleUpLimitWithScalingRules(args.CurrentReplicas, a.scaleUpEvents[args.Key], args.ScaleUpBehavior)
        ...
        // 扩容下限必须 <= hpaMaxReplicas
        maximumAllowedReplicas := args.MaxReplicas
        if maximumAllowedReplicas > scaleUpLimit {
            maximumAllowedReplicas = scaleUpLimit
            possibleLimitingReason = "ScaleUpLimit"
            possibleLimitingMessage = "the desired replica count is increasing faster than the maximum scale rate"
        } else {
            possibleLimitingReason = "TooManyReplicas"
            possibleLimitingMessage = "the desired replica count is more than the maximum replica count"
        }
        // 扩容正本数必须 <= 上一步计算的 desiredReplicas
        if args.DesiredReplicas > maximumAllowedReplicas {return maximumAllowedReplicas, possibleLimitingReason, possibleLimitingMessage}
    } else if args.DesiredReplicas < args.CurrentReplicas {     // 缩容
        // 依据 scaleDown.Policies 计算缩容下限
        scaleDownLimit := calculateScaleDownLimitWithBehaviors(args.CurrentReplicas, a.scaleDownEvents[args.Key], args.ScaleDownBehavior)
        ...
        // 缩容下限必须 >= hpaMinReplicas
        minimumAllowedReplicas := args.MinReplicas
        if minimumAllowedReplicas < scaleDownLimit {
            minimumAllowedReplicas = scaleDownLimit
            possibleLimitingReason = "ScaleDownLimit"
            possibleLimitingMessage = "the desired replica count is decreasing faster than the maximum scale rate"
        } else {
            possibleLimitingMessage = "the desired replica count is less than the minimum replica count"
            possibleLimitingReason = "TooFewReplicas"
        }
        // 缩容正本数必须 >= 上一步计算的 desiredReplicas
        if args.DesiredReplicas < minimumAllowedReplicas {return minimumAllowedReplicas, possibleLimitingReason, possibleLimitingMessage}
    }
    return args.DesiredReplicas, "DesiredWithinRange", "the desired count is within the acceptable range"
}

扩容下限的计算，根据 policy.percent/pods/period:

  • 能够存在多个 policy：由 scaleUp.SelectPolicy 决定是抉择这些 policy 的 max、min 还是 disable；
// pkg/controller/podautoscaler/horizontal.go
func calculateScaleUpLimitWithScalingRules(currentReplicas int32, scaleEvents []timestampedScaleEvent, scalingRules *autoscalingv2.HPAScalingRules) int32 {
    var result int32
    var proposed int32
    var selectPolicyFn func(int32, int32) int32
    if *scalingRules.SelectPolicy == autoscalingv2.DisabledPolicySelect {return currentReplicas // Scaling is disabled} else if *scalingRules.SelectPolicy == autoscalingv2.MinPolicySelect {
        result = math.MaxInt32
        selectPolicyFn = min // For scaling up, the lowest change ('min' policy) produces a minimum value
    } else {
        result = math.MinInt32
        selectPolicyFn = max // Use the default policy otherwise to produce a highest possible change
    }
    for _, policy := range scalingRules.Policies {replicasAddedInCurrentPeriod := getReplicasChangePerPeriod(policy.PeriodSeconds, scaleEvents)
        periodStartReplicas := currentReplicas - replicasAddedInCurrentPeriod
        if policy.Type == autoscalingv2.PodsScalingPolicy {            // Pods
            proposed = periodStartReplicas + policy.Value
        } else if policy.Type == autoscalingv2.PercentScalingPolicy {    // Percent
            // the proposal has to be rounded up because the proposed change might not increase the replica count causing the target to never scale up
            proposed = int32(math.Ceil(float64(periodStartReplicas) * (1 + float64(policy.Value)/100)))
        }
        result = selectPolicyFn(result, proposed)
    }
    return result
}

对于每个 scaleUp 的 policy：

  • 首先，计算过来 policy.periodSeconds 这段时间的伸缩总正本数；
  • 而后，计算 以后正本数 – 过来 policy.periodSconds 工夫内的伸缩正本数；

  • 后面两步的目标：

    • 是抹平 policy.periodSeconds 工夫内的伸缩正本数；
    • 即下一次伸缩间隔上一次伸缩的冷却工夫；

  • 最初：

    • 若 policy.Type=Pods，则再 + policy.Pods.Value 正本数；
    • 若 policy.Type=Percent，则再 * (1 + percent.Value)/100= 最终正本数；

policy.periodSeconds 工夫内的伸缩总正本数的计算：

// pkg/controller/podautoscaler/horizontal.go
func getReplicasChangePerPeriod(periodSeconds int32, scaleEvents []timestampedScaleEvent) int32 {period := time.Second * time.Duration(periodSeconds)
    cutoff := time.Now().Add(-period)
    var replicas int32
    for _, rec := range scaleEvents {if rec.timestamp.After(cutoff) {replicas += rec.replicaChange    // 汇总 periodSeconds 工夫内的伸缩正本总数，扩容 =+M，缩容 =-N}
    }
    return replicas
}

缩容下限的计算，根据 policy.percent/pods/period：

  • 与扩容下限的计算方法相似；

  • 区别在于：因为是缩容

    • perioldStartReplicas = 以后正本数 + 过来 periodSeoncds 内伸缩正本数；
    • 若 policy.Type=Pods，则再 – policy.Pods.Value 正本数；
    • 若 policy.Type=Percent，则再 * (1 – percent.Value)/100= 最终正本数；
// pkg/controller/podautoscaler/horizontal.go
func calculateScaleDownLimitWithBehaviors(currentReplicas int32, scaleEvents []timestampedScaleEvent, scalingRules *autoscalingv2.HPAScalingRules) int32 {
    var result int32
    var proposed int32
    var selectPolicyFn func(int32, int32) int32
    if *scalingRules.SelectPolicy == autoscalingv2.DisabledPolicySelect {return currentReplicas // Scaling is disabled} else if *scalingRules.SelectPolicy == autoscalingv2.MinPolicySelect {
        result = math.MinInt32
        selectPolicyFn = max // For scaling down, the lowest change ('min' policy) produces a maximum value
    } else {
        result = math.MaxInt32
        selectPolicyFn = min // Use the default policy otherwise to produce a highest possible change
    }
    for _, policy := range scalingRules.Policies {replicasDeletedInCurrentPeriod := getReplicasChangePerPeriod(policy.PeriodSeconds, scaleEvents)
        periodStartReplicas := currentReplicas + replicasDeletedInCurrentPeriod
        if policy.Type == autoscalingv2.PodsScalingPolicy {        // Pod
            proposed = periodStartReplicas - policy.Value
        } else if policy.Type == autoscalingv2.PercentScalingPolicy {    // Percent
            proposed = int32(float64(periodStartReplicas) * (1 - float64(policy.Value)/100))
        }
        result = selectPolicyFn(result, proposed)
    }
    return result
}

参考：

1.https://zhuanlan.zhihu.com/p/245208287

正文完
kubernetes
发表至： kubernetes
2023-03-25
 0
关于kubernetes:kubernetes-k8s-v1240-安装dashboard面板
关于kubernetes:图解-K8s-核心概念和术语
关于kubernetes:分享画的-Kubernetes-的源码架构图
关于kubernetes:在Kubernetesk8s中部署-jenkins