Google Kubernetes Engine introduced Pod snapshots, saving running workloads’ CPU and GPU memory for restoration on demand.

Google Kubernetes Engine (GKE) introduced Pod snapshots to save a workload’s running state, including CPU and GPU memory, and restore it on demand. The feature concerns workload-state preservation rather than a new processor or additional physical compute capacity.

Google reports that internal tests showed Pod snapshots could reduce AI inference start-up time by as much as 89%. The source does not supply the test workload or baseline, and this is a start-up measurement—not a claim of 89% lower inference latency, energy consumption, or overall cost.

The September update also describes built-in scale-to-zero support and Horizontal Pod Autoscaler (HPA) integration with an Autoscaling Metric. HPA can additionally use custom Prometheus Query Language (PromQL) metrics alongside standard metrics to trigger workload scaling.

A separate feature in the same update, the open-source GKE Agent Substrate, is described as supporting millions of sandboxes at 10 times the density of standard container runtimes, with sub-500-millisecond resume operations and more than 500 suspend/resume activations per second. These are separate Agent Substrate claims and should not be attributed to Pod snapshots; Google also describes kernel and network isolation for that runtime.