Yep, this exact behavior is managed by Google Cloud's Cluster Autoscaler—specifically its predictive autoscaling capability, which was available for GKE clusters around the 1.8.x version you're running.

Here's the breakdown:

• Regular Cluster Autoscaler only kicks in when there are pending Pods that can't be scheduled on existing nodes. But predictive autoscaling takes it a step further: it analyzes your cluster's historical load patterns (like recurring peak traffic times from the past few weeks) to proactively add nodes before the actual load hits.
• This is designed to avoid the lag time associated with spinning up new nodes (which can take a few minutes) during peak periods, ensuring your workloads get scheduled immediately when demand spikes. That's why you're seeing idle nodes pop up 20-30 minutes before your usual peak—this is the system prepping for incoming traffic.
• For your GKE 1.8.9-gke.1 cluster, this feature is integrated directly with the Cluster Autoscaler, not a separate tool. It might be enabled by default if your cluster shows consistent load patterns over time.

If you want to verify or adjust this behavior, you can:

• Run gcloud container clusters describe YOUR_CLUSTER_NAME --zone YOUR_ZONE and look for fields related to predictiveAutoscaling in the autoscaling configuration section.
• Check the GCP Console's cluster settings under "Autoscaling" to see if predictive mode is enabled, and tweak the settings (like the lead time for pre-provisioning nodes) if needed.

内容的提问来源于stack exchange，提问作者dimka