Updated on 2024-09-30 GMT+08:00

Creating a Node Scaling Policy

CCE provides auto scaling through the CCE Cluster Autoscaler add-on. Nodes with different flavors can be automatically added across AZs on demand.

If both a node scaling policy and the configuration in the Autoscaler add-on take effect, for example, there are pods that cannot be scheduled and the value of a metric reaches the threshold, scale-out is performed first for the unschedulable pods.

  • If the scale-out succeeds for the unschedulable pods, the system skips the metric-based rule logic and enters the next loop.
  • If the scale-out fails for the unschedulable pods, the metric-based rule is executed.

Prerequisites

Before using the node scaling function, you must install the CCE Cluster Autoscaler add-on of v1.13.8 or later in the cluster.

To use node flavor priorities, the Autoscaler version must be 1.19.35, 1.21.28, 1.23.30, 1.25.20, or later. To balance load among AZs, the version must be 1.23.122, 1.25.117, 1.27.85, 1.28.52, or later.

Notes and Constraints

  • If there are no nodes in a node pool, Autoscaler cannot obtain the CPU or memory data of the node, and the node scaling rule triggered using these metrics will not take effect.
  • If the driver of a GPU or NPU node is not installed, Autoscaler determines that the node is not fully available and the node scaling rules triggered using the CPU or memory metrics will not take effect.
  • When Autoscaler is used, some taints or annotations may affect auto scaling. Therefore, do not use the following taints or annotations in clusters:
    • ignore-taint.cluster-autoscaler.kubernetes.io: The taint works on nodes. Kubernetes-native Autoscaler supports protection against abnormal scale outs and periodically evaluates the proportion of available nodes in the cluster. When the proportion of non-ready nodes exceeds 45%, protection will be triggered. In this case, all nodes with the ignore-taint.cluster-autoscaler.kubernetes.io taint in the cluster are filtered out from the Autoscaler template and recorded as non-ready nodes, which affects cluster scaling.
    • cluster-autoscaler.kubernetes.io/enable-ds-eviction: The annotation works on pods, which determines whether DaemonSet pods can be evicted by Autoscaler. For details, see Well-Known Labels, Annotations and Taints.

Configuring Node Pool Scaling Policies

  1. Log in to the CCE console and click the cluster name to access the cluster console.
  2. In the navigation pane, choose Nodes. On the Node Pools tab, locate the row containing the target node pool and click Auto Scaling.

    • If Autoscaler has not been installed, configure add-on parameters based on service requirements, click Install, and wait until the add-on is installed. For details about add-on configurations, see CCE Cluster Autoscaler.
    • If Autoscaler has been installed, directly configure scaling policies.

  3. Configure auto scaling policies.

    AS Configuration

    • Customized Rule: Click Add Rule. In the dialog box displayed, configure parameters. You can add multiple node scaling policies, a maximum of one CPU usage-based rule, and one memory usage-based rule. The total number of rules cannot exceed 10.
      The following table lists custom rules.
      Table 1 Custom rules

      Rule Type

      Configuration

      Metric-based

      • Trigger: Select CPU allocation rate or Memory allocation rate and enter a value. The percentage must be greater than the value specified in the node resource requirements for a node scale-in when you configure a scaling policy (Configuring an Auto Scaling Policy for a Cluster).
        NOTE:
        • Resource allocation (%) = Resources requested by pods in the node pool/Resources allocatable to pods in the node pool
        • If multiple rules meet the conditions, the rules are executed in either of the following modes:

          If rules based on the CPU allocation rate and memory allocation rate are configured and two or more rules meet the scale-out conditions, the rule that will add the most nodes will be executed.

          If a rule is configured based on the CPU allocation rate and a periodic rule and both the rules meet the scale-out conditions, the periodic rule executed early changes the node pool to the scaling state. As a result, the metric-based rule cannot be executed. After the periodic rule is executed and the node pool status becomes normal, the metric-based rule will not be executed. If the metric-based rule is executed early, the periodic rule will be executed after the metric-based rule is executed.

        • If a rule is configured based on the CPU allocation rate and memory allocation rate, the policy detection period varies with the processing logic of each loop of the Autoscaler add-on. A scale-out is triggered once the conditions are met, but it is constrained by other factors such as the cooldown period and node pool status.
        • If the number of nodes reaches the upper limit of the cluster scale, the upper limit of the nodes supported in a node pool, or the upper limit of the nodes of a specific flavor, a metric-based scale-out will not be triggered.
        • If the number of nodes, CPUs, or memory resources reaches the upper limit for a node scale-out, a metric-based scale-out will not be triggered.
      • Action: Configure an action to be performed when the triggering condition is met.
        • Custom: Add a specified number of nodes to a node pool.
        • Auto calculation: When the trigger condition is met, nodes are automatically added and the allocation rate is restored to a value lower than the threshold. The formula is as follows:

          Number of nodes to be added = [Resource request of pods in the node pool/(Available resources of a single node x Target allocation rate)] – Number of current nodes + 1

      Periodic

      • Trigger Time: You can select a specific time every day, every week, every month, or every year.
      • Action: specifies an action to be carried out when the trigger time is reached. A specified number of nodes will be added to the node pool.
    • Nodes: The number of nodes in a node pool will always be within the range during auto scaling.
    • Cooldown Period: a period during which the nodes added in the current node pool cannot be scaled in.

    AS Object

    • Specifications: Configure whether to enable auto scaling for node flavors in a node pool.

      If multiple flavors are configured for a node pool, you can specify the upper limit for the number of nodes and the priority for each flavor separately.

  4. View cluster-level auto scaling configurations, which take effect for all node pools in the cluster. On this page, you can only view cluster-level auto scaling policies. To modify these policies, go to the Settings page. For details, see Configuring an Auto Scaling Policy for a Cluster.
  5. Click OK.

Configuring an Auto Scaling Policy for a Cluster

An auto scaling policy takes effect on all node pools in a cluster. After the policy is modified, the Autoscaler add-on will be restarted.

  1. Log in to the CCE console and click the cluster name to access the cluster console.
  2. In the navigation pane, choose Settings and click the Auto Scaling tab.

    • If Autoscaler has not been installed, configure add-on parameters based on service requirements, click Install, and wait until the add-on is installed. For details about add-on configurations, see CCE Cluster Autoscaler.
    • If Autoscaler has been installed, directly configure scaling policies.

  3. Configure for an elastic scale-out.

    • Auto Scale-out when the load cannot be scheduled: When workload pods in a cluster cannot be scheduled (pods remain in pending state), CCE automatically adds nodes to the slave node pool. If a pod has been scheduled to a node, the node will not be involved in an automatic scale-out. Such auto scaling typically works with an HPA policy. For details, see Using HPA and CA for Auto Scaling of Workloads and Nodes.

      If this function is not enabled, custom scaling rules are the only option for performing a scale-out.

    • Upper limit of resources to be expanded: the upper limit for the cluster's resources, such as the number of nodes, CPU cores, and memory. Once this limit is reached, no new nodes will be automatically added.
    • Scale-Out Priority: You can drag and drop the node pools in a list to adjust their scale-out priorities.

  4. Configure for an elastic scale-in. Elastic scale-in is disabled by default. After it is enabled, the following configurations are supported:

    Node Scale-In Conditions: Nodes in a cluster are automatically scaled in when the scale-in conditions are met.
    • Node Resource Condition: When the requested cluster node resources (both CPU and memory) are lower than a certain percentage (50% by default) for a period of time (10 minutes by default), a cluster scale-in is triggered.
    • Node Status Condition: If a node is unavailable for a specified period of time, the node will be automatically reclaimed. The default value is 20 minutes.
    • Scale-in Exception Scenarios: When a node meets the following exception scenarios, CCE will not scale in the node even if the node resources or status meets scale-in conditions:
      1. Resources on other nodes in the cluster are insufficient.
      2. Scale-in protection is enabled on the node. To enable or disable node scale-in protection, choose Nodes in the navigation pane and then click the Nodes tab. Locate the target node, choose More, and then enable or disable node scale-in protection in the Operation column.
      3. There is a pod with the non-scale label on the node.
      4. Policies such as reliability have been configured on some containers on the node.
      5. There are non-DaemonSet containers in the kube-system namespace on the node.
      6. (Optional) A container managed by a third-party pod controller is running on a node. Third-party pod controllers are for custom workloads except Kubernetes-native workloads such as Deployments and StatefulSets. Such controllers can be created using CustomResourceDefinitions.
    Node Scale-in Policy
    • Number of Concurrent Scale-In Requests: maximum number of idle nodes that can be concurrently deleted. Default value: 10.
      Only idle nodes can be concurrently scaled in. Nodes that are not idle can only be scaled in one by one.

      During a node scale-in, if the pods on the node do not need to be evicted (such as DaemonSet pods), the node is idle. Otherwise, the node is not idle.

    • Node Recheck Timeout: interval for rechecking a node that could not be removed. Default value: 5 minutes.
    • Cooldown Time
      • Scale-in Cooldown Time After Node Deletion: cooldown period for starting scale-in evaluation again after an auto scale-in is triggered in a cluster. Default value: 10 minutes.
      • Scale-in Cooldown Time After Scale-out: cooldown period for starting scale-in evaluation again after an auto scale-out is triggered in a cluster. Default value: 10 minutes.

        If both auto scale-out and scale-in exist in a cluster, set Scale-in Cooldown Time After Scale-out to 0 minutes. This prevents the node scale-in from being blocked due to continuous scale-out of some node pools or retries upon a scale-out failure, which results in unexpected waste of node resources.

      • Scale-in Cooldown Time After Failure: cooldown period for starting scale-in evaluation again after an auto scale-in is failed in a cluster. Default value: 3 minutes. For details, see Cooldown Period.

  5. Click Confirm configuration.

Cooldown Period

The impact and relationship between the two cooldown periods configured for a node pool are as follows:

Cooldown Period During a Scale-out

This interval indicates the period during which nodes added to the current node pool after a scale-out cannot be deleted. This setting takes effect in the entire node pool.

Cooldown Period During a Scale-in

The interval after a scale-out indicates the period during which the entire cluster cannot be scaled in after the Autoscaler add-on triggers a scale-out (due to the unschedulable pods, metrics, and scaling policies). This interval takes effect in the entire cluster.

The interval after a node is deleted indicates the period during which the cluster cannot be scaled in after the Autoscaler add-on triggers a scale-in. This setting takes effect in the entire cluster.

The interval after a failed scale-in indicates the period during which the cluster cannot be scaled in after the Autoscaler add-on triggers a scale-in. This setting takes effect in the entire cluster.

Period for Autoscaler to Retry a Scale-out

If a node pool failed to scale out, for example, due to insufficient resources or quota, or an error occurred during node installation, Autoscaler can retry the scale-out in the node pool or switch to another node pool. The retry period varies depending on failure causes:

  • When resources in a node pool are sold out or the user quota is insufficient, Autoscaler cools down the node pool for 5 minutes, 10 minutes, or 20 minutes. The maximum cooldown duration is 30 minutes. Then, Autoscaler switches to another node pool for a scale-out in the next 10 seconds until the expected node is added or all node pools are cooled down.
  • If an error occurred during node installation in a node pool, the node pool enters a 5-minute cooldown period. After the period expires, Autoscaler can trigger a node pool scale-out again. If the faulty node is automatically reclaimed, Cluster Autoscaler re-evaluates the cluster status within 1 minute and triggers a node pool scale-out as needed.
  • During a node pool scale-out, if a node remains in the installing state for a long time, Cluster Autoscaler tolerates the node for a maximum of 15 minutes. After the tolerance period expires, Cluster Autoscaler re-evaluates the cluster status and triggers a node pool scale-out as needed.

Example YAML

The following is a YAML example of a node scaling policy:

apiVersion: autoscaling.cce.io/v1alpha1
kind: HorizontalNodeAutoscaler
metadata:
  name: xxxx
  namespace: kube-system
spec:
  disable: false
  rules:
  - action:
      type: ScaleUp
      unit: Node
      value: 1
    cronTrigger:
      schedule: 47 20 * * *
    disable: false
    ruleName: cronrule
    type: Cron
  - action:
      type: ScaleUp
      unit: Node
      value: 2
    disable: false
    metricTrigger:
      metricName: Cpu
      metricOperation: '>'
      metricValue: "40"
      unit: Percent
    ruleName: metricrule
    type: Metric
  targetNodepoolIds:
  - 7d48eca7-3419-11ea-bc29-0255ac1001a8
Table 2 Key parameters

Parameter

Type

Description

spec.disable

Bool

Whether to enable the scaling policy. This parameter takes effect for all rules in the policy.

spec.rules

Array

All rules in a scaling policy.

spec.rules[x].ruleName

String

Rule name.

spec.rules[x].type

String

Rule type. Cron and Metric are supported.

spec.rules[x].disable

Bool

Rule switch. Currently, only false is supported.

spec.rules[x].action.type

String

Rule action type. Currently, only ScaleUp is supported.

spec.rules[x].action.unit

String

Rule action unit. Currently, only Node is supported.

spec.rules[x].action.value

Integer

Rule action value.

spec.rules[x].cronTrigger

N/A

Optional. This parameter is valid only in periodic rules.

spec.rules[x].cronTrigger.schedule

String

Cron expression of a periodic rule.

spec.rules[x].metricTrigger

N/A

Optional. This parameter is valid only in metric-based rules.

spec.rules[x].metricTrigger.metricName

String

Metric of a metric-based rule. Currently, Cpu and Memory are supported.

spec.rules[x].metricTrigger.metricOperation

String

Comparison operator of a metric-based rule. Currently, only > is supported.

spec.rules[x].metricTrigger.metricValue

String

Threshold of the metric rule. The value can be an integer ranging from 1 to 100 and must be a character. If the value is set to -1, the threshold is automatically calculated.

spec.rules[x].metricTrigger.Unit

String

Unit of the metric-based rule threshold. Currently, only % is supported.

spec.targetNodepoolIds

Array

All node pools associated with the scaling policy.

spec.targetNodepoolIds[x]

String

UID of the node pool associated with the scaling policy.