SIG Autoscaling Updates and Feature Highlights
Conference: KubeCon + CloudNativeCon Europe 2023
2023-04-20
Authors: Guy Templeton, Chen Wang, Michele Orlandi, Piotr Betkier, Jayant Jain
Summary
Introduction of Multi-Dimensional Pod Auto Scaler for Kubernetes
- Current auto scaling controllers available in the community are Horizontal Pod Auto Scaler (HPA) and Vertical Pod Auto Scaler (VPA)
- Multi-Dimensional Pod Auto Scaler (MPA) is a new in-house proposal for Kubernetes auto scaler
- MPA consists of three main controllers: Recommended, Updater, and Admission Controller
- Recommended analyzes historical usage patterns of the Pod and recommends resource usage and limit to be set based on the histogram of the resource usage observed in the previous time window
- Updater observes the difference between the recommended requests and limits and the current set request and limit and will evict the parts if the gap is too big
- Admission Controller is responsible for updating the pause request a limit when the past the victim pass are restarting according to the recommended values provided by the recommended
- MPA provides automatic vertical scaling for Kubernetes pod allowing them to adjust resource requests and limits based on actual usage
Abstract
In this talk the SIG will be covering new features as well as proposed future functionality. New features including the Cluster Autoscaler's snapshotter to assist in cluster debugging. This allows cluster operators to understand the state the CA is using to drive decisions. Proposed future functionality includes: The Multidimensional Pod Autoscaler (MPA): The MPA proposes combining the actions of the Horizontal and Vertical Pod Autoscalers, but separating the actuation from the controlling algorithms. The Balancer API: Balancer is a new resource in the Autoscaling API that may be used as an HPA scale target. It allows autoscaling workloads in which some Pods have a slightly different spec. Use cases include evenly distributing Pods across zones in regional clusters or maintaining a desired ratio between Pods running on differing VMs. Promotion of a number of alpha HPA features. If you're interested in any of these, or want to be involved in implementing them, come along!
Materials:
Post a comment
Related work
Conference: KubeCon + CloudNativeCon Europe 2022
Authors: Michael McCune, Guy Templeton, David Morrison, Joachim Bartosik
2022-05-19
Conference: KubeCon + CloudNativeCon North America 2021
Authors: Marcin Wielgus, Joseph Burnett
2021-10-14
Conference: KubeCon + CloudNativeCon North America 2022
Authors: Marcin Wielgus, Diego Bonfigli, Jayant Jain
2022-10-28