GCP Idle Findings: Min Instances, Spanner, Bigtable, LBs
The GCP idle rules look for provisioned capacity with no callers: warm instances, database instances and forwarding rules that bill while nothing uses them.
Several GCP rules target capacity provisioned ahead of demand, such as Cloud Run minimum instances, which Google bills while they are kept warm. The GCP idle rules read each service’s own Cloud Monitoring metric by name and fire only when it shows zero measured use over the window.
Warm instances on Cloud Run
GCP Cloud Run Service With Idle Min Instances is the gentlest idle lever ZopNight has. A service with minimum instances above zero and no requests for 30 days is recommended to scale its minimum to zero. It is reversible: Cloud Run bills instances kept warm by the minimum-instances setting, while at a minimum of zero the service scales down when there is no traffic and pays a cold start on the next request instead. The saving is priced from the warm vCPU and memory the minimum holds. See the Cloud Run fact sheet and GCP Cloud Run.
Databases and caches with no requests
- GCP Cloud Spanner Instance Idle fires when the instance’s request count is zero over the window.
- GCP Bigtable Instance Idle requires zero requests and uses CPU load as a veto, so an instance still doing internal work is left alone.
- GCP Memorystore Instance Idle requires zero client connections, with operations per second as a corroborating check. See Memorystore and GCP Memorystore.
Each prices the full monthly run-rate of the instance, because an instance with no callers is recoverable in full.
Load balancers
GKE Idle Load Balancer and GCP Idle Global Load Balancer fire only when every traffic axis the rule can read, including requests, new connections and ingress and egress bytes, sits below a small floor on both average and peak for 30 days. Egress is included so that a balancer whose traffic is mostly outbound is not mistaken for idle.
Data processing and ML
GCP Dataproc Cluster Idle prices the whole cluster. GCP Vertex AI Endpoint Idle fires on zero predictions; when schedule data exists it prefers a recurring off-hours window over removing the endpoint. Feature store and vector search endpoint rules cover the remaining Vertex AI serving capacity.
What is not here
An idle Compute Engine VM does not appear in this category. On GCP, quiet VMs are routed to schedule recommendations, where the fix is an off-hours window, and stopped VMs with leftover disks are covered by GCP orphans.
Acting
Dataproc is on ZopNight’s pause roster, and that pause is terminal: there is no restore step. Other findings link to the console with numbered steps. Remediation needs a read + write project connection. Full list: GCP recommendations.