# Scheduling GKE Node Pool

> A node pool schedule scales just that pool to 0 and reinstates the recorded size on resume. Whether money is actually saved depends on the pods: workloads pinned by selectors go Pending, but unpinned pods reschedule onto the other pools, where they keep running and billing on different nodes.

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

---

## Zero for one pool, business as usual elsewhere

The schedule targets a single pool (commonly the expensive one, GPU or high-memory), scales
it to 0 nodes and records the previous size for the resume. The rest of the cluster keeps its
nodes, its add-ons and its workloads, which is what makes single-pool scheduling attractive:
the blast radius is one team's hardware, not the shared cluster.

## Pods do not always go quiet, some just move

Eviction gives every displaced pod a choice made by the scheduler. Pods bound to the pool by
node selectors, taints, tolerations or GPU resource requests have nowhere to go and wait as
Pending, costing nothing. Pods that merely happened to be on the pool are rescheduled onto
whatever else has room. They keep running, and their consumption shifts onto the surviving
pools, potentially triggering scale-up there. The saving is real only for the pinned
workloads; measure cluster-wide node spend before and after, not the scheduled pool alone.

## The autoscaler will refill what pods demand

A pool under cluster-autoscaler management scales back up when Pending pods request its
resources. If anything overnight creates pods pinned to the scheduled pool, say a CronJob
requesting a GPU at 1am, the autoscaler obliges and the pool quietly runs all night. Either
exclude the pool from autoscaling during the window or move those jobs inside business hours.

## Resume restores the recorded size

The pool returns to the count captured at stop time, not to whatever someone set in between.
Nodes then boot, join and pass readiness before the Pending backlog schedules. On GPU pools
with driver installation, that is minutes, so lead the first training job accordingly.

## The natural fit

GPU pools used interactively by researchers are the canonical case. Pinned workloads, diurnal
usage and expensive nodes are every property this schedule wants, with 12-14 idle hours a night
to harvest.