Metric-driven autoscaling
Change replica count within reviewed bounds without treating missing metrics as idle capacity.
Autoscaling reconciles a managed droplet group toward a replica count calculated from an approved metric policy. It creates or removes group members; it does not resize arbitrary existing guests.
Define the group
Open Compute → Autoscaling groups → Create group in the target project. Edit bounds, policy, and paused state on that group's Configuration tab.
Select a pinned template revision, region, network, placement group, minimum replicas, and maximum replicas. All replicas share the declared managed group identity. The policy must have quota and eligible capacity for increases.
Scale from zero requires a supported metric source independent of the missing guests. A per-guest CPU average cannot establish demand when no guest exists.
Evaluate demand
Choose metric, aggregation, target, evaluation window, and stabilization duration. Missing or stale observations leave evaluation unknown and prevent a destructive scale-down. A threshold increase does not bypass image or placement validation.
The group records desired replicas, ready replicas, pending creates, and pending removals. Count in-flight creates when evaluating another increase so slow boot does not produce an uncontrolled sequence of duplicate requests.
Scale out
New guests use the group's pinned template revision. They enter service only after configured initialization and readiness checks. A create failure records its resources and retry condition; it does not advance ready count.
Capacity or spend admission can block an increase below the configured maximum. The status identifies the limiting condition rather than reporting the policy as satisfied because it is enabled.
Scale in
The policy selects only owned group members. It respects minimum replicas, protection, disruption constraints, and required backend drain. A member with exclusive state or an unresolved operation can block removal.
Scaling is unsuitable for a database primary without an application-specific membership and promotion procedure. Backup availability does not make arbitrary member termination safe.
Pause and override
Pause stops new reconciliation actions while in-flight operations resolve. A manual desired count can be overwritten after automatic evaluation resumes; record whether an override is temporary or a policy revision.
Changing the template affects new members and does not silently replace the current group. Use rolling replacement for version rollout. See Autoscaling API and spend admission.
A policy can use an approved fleet-wide aggregate when its controller has permission to read that source. The group still changes only its owned droplet members; it does not provision new physical hosts to resolve a capacity shortage.