Google Cloud has introduced support for concurrent node pool auto-creation in its Google Kubernetes Engine (GKE), a change designed to reduce cluster provisioning time and improve autoscaling performance.

Node pools, a group of virtual machines (VMs) with identical hardware and configuration, allow Kubernetes clusters to run workloads with different resource requirements – such as CPUs, GPUs, TPUs, or pricing models – in the same cluster. The update targets the creation of multiple node pools simultaneously, rather than sequentially, to tackle a long-standing bottleneck in cluster scale-out operations.

According to Google, internal benchmarks show provisioning speeds improving by up to 85% in some scenarios. The change primarily affects clusters that require multiple node pools, such as those running heterogeneous workloads, multi-tenant applications, or large AI training jobs.

The issue of compounded cluster delays

Until now, node pool creation was handled one operation at a time. Since creating a new node pool typically takes 30 to 45 seconds, clusters that needed several node pools faced compounded delays during deployment and scaling events.

This sequential process reduced autoscaling responsiveness, particularly in clusters where multiple node pools were required at once.

This prompted Google Cloud's team to add the new concurrency support so GKE can perform multiple node pool creation operations in parallel. This allows clusters to scale out to different node types quicker and reduce the time workloads spend waiting for infrastructure to become available.

Google highlighted several scenarios where parallel node pool creation provides the most benefit, including heterogeneous and multi-tenant clusters, where different workloads or teams require distinct node configurations; and AI and machine learning workloads, including multi-host TPU training jobs, which often require separate node pools for each hardware slice.

In addition, parallel node pool creation is said to contribute to cost-optimized deployments, such as those using Google's preemptibleSpot VM instances or multiple ComputeClass priority levels, each of which must be isolated into separate node pools.