CPA Introduction

Welcome! In this guide, we cover the Cluster Proportional Autoscaler (CPA)—a Kubernetes component that automatically scales auxiliary services (DNS, network controllers, metrics servers) in proportion to your cluster’s worker nodes or CPU cores. CPA is different from the Cluster Autoscaler; it adjusts supporting deployments instead of core node counts.

How CPA Works

CPA continuously watches the size of your cluster—either the number of schedulable worker nodes or the total CPU cores—and computes desired replicas for each target deployment using a simple scaling formula.

After calculating the target replica count, CPA calls the Kubernetes API to patch the deployment spec and keep your supporting services in sync.

Scaling Modes

CPA supports two scaling modes—ladder and linear—each suited to different workload patterns:

• Ladder Mode: Uses discrete thresholds for stepwise scaling—great for predictable, SLA-driven workloads.
• Linear Mode: Applies a continuous ratio between replicas and cluster size—ideal for elastic, usage-based scaling.

Ladder Mode Configuration

Define one or more thresholds to map CPU cores or nodes to replica counts:

data:
  ladder: |-
    {
      "coresToReplicas": [[64,3]],
      "nodesToReplicas": [],
      "includeUnschedulableNodes": false
    }

Add node-based scaling:

data:
  ladder: |-
    {
      "coresToReplicas": [[64,3]],
      "nodesToReplicas": [[2,2]],
      "includeUnschedulableNodes": false
    }

Full ladder example with multiple steps:

data:
  ladder: |-
    {
      "coresToReplicas": [
        [1,      1],
        [64,     3],
        [512,    5]
      ],
      "nodesToReplicas": [
        [1,      1],
        [2,      2]
      ],
      "includeUnschedulableNodes": false
    }

In this configuration:

• 1–63 cores → 1 replica
• 64–511 cores → 3 replicas
• ≥512 cores → 5 replicas
• 1 node → 1 replica; ≥2 nodes → 2 replicas

Linear Mode Configuration

Linear scaling defines an exact ratio of cluster resources to replicas, along with bounds:

data:
  linear: |-
    {
      "coresPerReplica": 2,
      "nodesPerReplica": 1,
      "min": 1,
      "max": 100,
      "preventSinglePointFailure": true,
      "includeUnschedulableNodes": true
    }

Replica Calculation

CPA computes replicas in three steps:

replicas = max(
  ceil(cores / coresPerReplica),
  ceil(nodes / nodesPerReplica)
)
replicas = min(replicas, max)
replicas = max(replicas, min)

For a cluster with 4 nodes and 13 cores:

ceil(13 / 2) = 7   # coresPerReplica = 2
ceil(4  / 1) = 4   # nodesPerReplica = 1

replicas = max(7, 4) = 7
replicas = min(7, 100) = 7
replicas = max(7, 1)   = 7

CPA will deploy 7 replicas to meet the defined ratios and bounds.

Summary

• Ladder Mode: Stepwise thresholds for controlled scaling.
• Linear Mode: Continuous scaling with direct resource-to-replica ratios.
• CPA integrates with the Kubernetes API to automatically adjust replicas of auxiliary services based on cluster size.

References and Further Reading

• CPA GitHub Repository
• Kubernetes Autoscaling Concepts
• Cluster Autoscaler