Cluster Autoscaler Lab

Overview

Cluster Autoscaler Lab is a project for testing how pending pods, node groups, taints, requests, and scheduling constraints affect autoscaling decisions.

Motivation

Autoscaling is often treated as a checkbox, but the behavior depends on workload requests, node capacity, cloud provider limits, and scheduler signals. This lab makes those interactions visible.

Architecture

flowchart LR
  PendingPods[Pending Pods] --> Scheduler[Scheduler]
  Scheduler --> Autoscaler[Cluster Autoscaler]
  Autoscaler --> Cloud[AWS Auto Scaling Group]
  Cloud --> Nodes[Worker Nodes]

Design decisions

• Use Terraform to keep the AWS environment reproducible.
• Keep workload fixtures small and easy to change.
• Capture scheduler and autoscaler logs for each scenario.
• Test both scale-up and scale-down behavior.

Challenges

Autoscaling bugs are frequently configuration bugs. Resource requests, topology constraints, taints, and max node group size can all produce similar symptoms.

Lessons learned

The fastest way to debug autoscaling is to compare what the scheduler cannot place with what the autoscaler is allowed to create.

Screenshots

Future improvements

• Add scenario-based runbooks.
• Compare managed node groups and self-managed groups.
• Add Grafana dashboard examples.
• Publish Terraform modules for the lab.