README.md

Documentation

Contents

• Best Practices, Common Issues and Gotchas
  • Best Practices
  • Common Issues and Gotchas
• Configuration
  • Command line options
  • Node Group Configuration
  • Advanced Configuration
• Deployment
  • AWS
• Algorithm and Calculations
  • Assumptions
  • Requests, capacity and utilisation
  • Scale up delta
• Scale Process
  • Scale up
  • Scale down
  • Scale lock
  • Tainting of nodes
  • Cordoning of nodes
• Node Termination
  • Node selection method for termination
  • Annotating nodes / Stopping termination of selected nodes
• Pod and Node Selectors
  • Nodes
  • Pods
  • Default Pod Selector
  • More Information
• Metrics
  • Exposed Metrics
  • Grafana
  • Recommendations
• Glossary

Package Layout and Usage

• cmd
  • contains command function, setup, and config loading
• pkg/controller
  • contains the core logic specific to escalator and nodegroups
• pkg/k8s
  • provides application utils and help with interfacing with Kubernetes and client-go
• pkg/cloudprovider
  • provides everything related to cloud providers
  • pkg/cloudprovider/aws
    • provides the aws implementation of cloudprovider
• pkg/metrics
  • provides a place for all metric setup to live
• pkg/test
  • provides Kubernetes and cloudprovider helpers for testing

Design

Summary

Full scale up/down process can be found here

• The autoscaler is designed as a static autoscaler.
• Start watching all pods and nodes in the cluster for a certain node group, including pending pods
• Every tick the autoscaler will check the overall cluster utilisation for each node group and make a decision on whether to do nothing, scale up, or scale down:
  • If the scale lock is present
    • Check whether the scale lock has timed out, otherwise do nothing
  • On a scale up event
    • Untaint as many nodes as we need to
    • Increase the ASG size by the remaining nodes
    • Lock the scale lock until ASG resize is complete to prevent infinite scaling
  • On a scale down event
    • Taint the configured amount of nodes so they are unschedulable
    • Wait until the grace period for the node has passed and the node is empty
    • Terminate any nodes that have passed the grace period and are empty
  • On a "do nothing" event
    • Terminate any nodes that have passed the grace period and are empty