Autoscaling is a widely adopted feature of CCSE. Typical scenarios include online business elasticity, large-scale computing training, deep learning GPU or shared GPU training and inference, and cron periodic load changes. Autoscaling has two dimensions:
Scheduling layer elasticity, which is primarily responsible for adjusting the scheduling capacity of workloads. A typical component for scheduling layer elasticity is the Horizontal Pod Autoscaler (HPA), which adjusts the number of application replicas to change the scheduling capacity of the current load, achieving elasticity at the scheduling layer.
Resource layer elasticity, which mainly pertains to when the capacity planning of the cluster cannot meet its scheduling capacity. In this case, resources such as ECS instances are scaled out to supplement the scheduling capacity.
These two layers of elastic components and capabilities can be used independently or together, and they are decoupled through the capacity status of the scheduling layer.