# Scheduling GKE Cluster

> GKE cluster schedules scale every node pool to 0, saving each pool's size and restoring it at start. The control plane keeps running throughout, so kubectl and GitOps work all night against a cluster with no nodes. The management fee of $0.10 per cluster-hour continues regardless.

Source: https://zop.dev/integrations/gcp/gke-cluster/schedule
Updated: 2026-08-19

---

## Nodes leave, the control plane stays

Scaling all node pools to 0 empties the cluster of compute while leaving the API server
reachable. Every workload sits Pending, but kubectl responds, manifests apply, and GitOps
controllers reconcile all night: the reconciled state simply has nowhere to run until nodes
return. Overnight CI that only applies manifests keeps passing; anything that waits for a
rollout to become ready will hang until morning.

## The fee that ignores the schedule

GKE charges a flat management fee of $0.10 per cluster per hour, roughly $73 a month, no
matter how many nodes exist, with the free tier covering one zonal or Autopilot cluster per
account. A scheduled-to-zero cluster keeps paying it. For a fleet of small dev clusters the
fee becomes the dominant residual, which is a quiet argument for fewer, shared dev clusters
rather than one per team.

Autopilot clusters sit outside this mechanism entirely. They bill per pod rather than per node,
and have no user-managed pools to scale.

## Per-pool memory, per-pool restore

Each pool's node count is recorded at stop and reinstated at start, pool by pool. A cluster
with a system pool of 3 and a GPU pool of 2 comes back as exactly that. Interim manual edits
made while the pools sit at zero are overwritten by the recorded sizes.

## System workloads go Pending too

Emptying every pool evicts cluster add-ons along with application pods. They recover unaided
when nodes return, but the morning has an ordering: nodes register, system pods land first,
then the application backlog places. Allow 5-10 minutes from the start cron before
health checks judge the result.

## Automation that refills pools

Node auto-provisioning and pool-level autoscalers respond to Pending pods, and a
scaled-to-zero cluster is nothing but Pending pods. Left enabled overnight, they will
faithfully undo the schedule. Scheduled clusters need those features scoped, disabled at
night, or hosting only workloads that scale to zero themselves.