Skip to main content

Control Costs

Two actions cut cost.

  1. Size each resource with the estimate in front of you. Each pull request shows a monthly estimate for the fixed-cost parts it can price. The estimate is not the bill. Some resources have no estimate. Requests, traffic, storage and AI use cost more on top.
  2. Put non-production projects on a hibernation schedule, so that they sleep when nobody works.

What hibernation switches off​

Hibernation removes compute. It does not delete data.

ResourceWhat hibernation doesBilling effect
AlloyDBRemoves the instances. The cluster, storage and backups stay.Instance billing stops. Storage and backups keep billing.
Cloud SpannerScales to the 100-processing-unit floor and turns off autoscaling.Falls to the minimum. It does not reach zero.
Applications on Compute EngineSets the instance group to zero.Compute billing stops.
Applications on KubernetesSets replicas to zero.Pods stop. The cluster keeps running.
Virtual machinesSets the scaling floor and ceiling to zero.Compute billing stops.
Public ingress gatewaysSets the gateway group to zero.The environment is unreachable while it sleeps.
Applications on Cloud RunNothing. The scaling block of the DeploymentConfig sets Cloud Run scaling.No change.
Redis, GKE clusters, MCP and Agent deploymentsNothing today.These keep running and keep billing.

AlloyDB has no pause operation. The engine removes the instances and keeps the cluster. When the project wakes, AlloyDB takes several minutes to return. Set the wake-up time well before the first user needs the database.

warning

Use hibernation for development, integration, staging and sandbox projects only. The platform lets you add a schedule to a production project. Make "never on production" a review rule.

Declare a schedule​

Hibernation is a field of the Project manifest. A window has a start and an end, both cron expressions. Write both. A window with one boundary fails at plan time.

apiVersion: lowops.manifests.v1
kind: Project
metadata:
name: my-project
environment: development
organizational-unit: my-unit
organization: my-org
spec:
hibernation:
hibernate: false # true sleeps the project now and ignores all windows
windows:
overnight: # your own name for the window
start: "0 20 * * *" # asleep from 20:00 UTC
end: "0 7 * * *" # awake at 07:00 UTC
  • Time zone. The cron expressions use UTC. A window has no time-zone field, so convert your office hours to UTC. If your region changes its clocks, edit the manifest each time.

  • Many windows. The project sleeps if any window says so. Use a second window for weekends.

  • Exclusions. An exclusion stops the schedule for a period, for example a release week. It uses timestamps with a UTC offset.

    spec:
    hibernation:
    exclusions:
    release-week:
    start: "2026-10-05T00:00:00Z"
    end: "2026-10-09T23:59:59Z"

    An exclusion does not schedule its own wake-up. If it starts when no window boundary fires, the environment stays asleep until the next boundary or the next merged change. Start an exclusion before a window boundary.

How the schedule runs​

Each different cron boundary in your organization is one scheduled trigger. When it fires, it runs a full reconcile of the organization against main. Two results follow:

  • main must always be in a state that can apply. A boundary applies everything on main, not only the hibernation change.
  • Projects with the same boundary share one trigger. Reuse the same times across projects.

Inheritance​

You can declare hibernation on the organization, the OU, the environment or the project. A parent that hibernates puts all its children to sleep. A child cannot opt out with hibernate: false. Declare hibernation on single non-production projects. Move it to a higher level only when you mean every project below it.

Right-size​

  • Databases. spec.alloydb.cpuCount is the largest number on a database bill. The floor is 1 vCPU. If a region rejects 1 vCPU, 2 is the floor there.
  • Read replicas. A clusterSize above 1 adds read-pool instances. Each bills like the primary. Non-production rarely needs them.
  • Cloud Run. Keep a warm minimum only where a cold start is not acceptable, such as a login path.
  • Compute target. A move from Kubernetes to Cloud Run is a one-field change. See Migrate a Service to Cloud Run.

Next​