Implementing Second-Level Elastic Scaling with cce-autoscaling-placeholder

Component introduction

After enabling auto-scaling for a CCE node group, if Pod scheduling fails due to insufficient resources, the cluster automatically scales up nodes. This typically takes a few minutes. However, during traffic surges, this minute-level scaling may not meet service demands quickly enough. This document outlines how to use K8S PriorityClass to create Pod placeholders, enabling near-instant scaling in CCE for such scenarios.

Implementation principle

The cce-autoscaling-placeholder uses low-priority Pods to preoccupy resources and reserve some resources as a buffer. When Pod is scaled, high-priority Pods can quickly preempt resources from low-priority Pods for scheduling. The low-priority cce-autoscaling-placeholder Pods will be "preempted" and enter the pending status. If a node group is configured and auto scaling is enabled, this will trigger node scale-up. For more information, refer to Node Group Management.

By reserving some resources as buffers, even with slow node scaling, certain Pods can be quickly scaled and scheduled, achieving near-instant scaling. To adjust the buffer resource amount, modify the request or replica count of the cce-autoscaling-placeholder based on actual needs.

Operation steps

1. Sign in to the Baidu AI Cloud official website and enter the management console.
2. Select Product Tour - Containers - Baidu Container Engine and click to enter the CCE management console.
3. Click Helm > Helm Templates in the left navigation bar.
4. In Baidu AI Cloud Templates, select the cce-autoscaling-placeholder template, and click Install to deploy it.
5. Complete the configurations detailed in the template.

6. Click OK button to complete the deployment.
7. When creating a Nginx Pod, you will notice that the Nginx Pod can quickly preempt Placeholder Pod for quick startup, while the Placeholder will trigger a new node scale-up, as shown below:

Reference

• Pod Priority and Preemption