Google Kubernetes Engine 1.37 adds native scaling to and from zero replicas, restarting workloads on capacity buffers when demand returns.

Google announced native scale-to-zero functionality in Google Kubernetes Engine (GKE) 1.37 for intermittently active workloads such as batch processors, event-driven workers, and development environments. Workloads can scale down to zero replicas and restart on GKE capacity buffers when demand returns; this does not mean the entire cluster or its reserved capacity scales to zero.

The implementation combines the Horizontal Pod Autoscaler (HPA) with AutoscalingMetric, Google’s managed metrics signal pipeline. It builds on Kubernetes Enhancement Proposal 2021 (KEP-2021) to support `minReplicas: 0`, stopping pods below a metric threshold and restarting the deployment when a signal indicates pending work.

Google contrasts this native control-plane functionality with Kubernetes Event-Driven Autoscaling (KEDA), an optional component that requires operators and additional configuration. Its example uses an AutoscalingMetric custom resource definition (CRD) and an HPA to scale a worker from the number of undelivered messages in a Pub/Sub subscription.

A related announcement adds native Prometheus Query Language (PromQL) support for metrics in Cloud Monitoring through Google Managed Service for Prometheus, without third-party adapters. Google’s cited 5-second metrics-reading latency concerns its earlier native custom-metrics pipeline—not a guaranteed workload cold-start time—and self-hosted Prometheus support remains planned for general availability after preview.