Skip to content

RB002: Range teardown stuck

Applicability: GCP/GCE primary; AWS optional. Operator: range and security operator. Verified: 2026-10-23.

Symptom

A range remains destroying, a cancel or expiry does not converge, or the portal shows destroyed while provider resources are still present.

Investigation

  1. Record request ID, destroy operation ID, generation, persisted backend binding, cancellation state, lease, and cleanup-verification state.
  2. Inspect worker-provisioner-launcher, provisioner Jobs, the operation-result applier, and worker-reconciler for the bounded incident window.
  3. Inventory GCE resources by ownership labels. Include instances, disks, subnets, routes/NAT enrollment, addresses, firewall bindings, and dynamic secrets. For AWS, include EC2, ENIs, volumes, security groups, routes, and provisioner state.
  4. Treat a logical destroyed status and a successful Job exit as separate evidence from an empty provider inventory.

Resolution

If the application still offers cancel or destroy, invoke it once through the authorized lifecycle service and follow the resulting operation ID. Do not delete provider resources first: that breaks the operation-result and cleanup proof. If a credential must be revoked to contain an active range, use the approved credential/lifecycle service, preserve the accepted cleanup authority, and confirm that the destroy intent remains executable after caller revocation.

When the owning operation is terminal but residuals remain, block reuse of its network and identifiers and escalate to the platform owner. GCP has no approved general orphan-mutation command. Verify completion only after cleanup verification and independent provider inventory agree.

Prevention

Monitor destroy-operation age, interrupt backlog, cleanup-verification outcome, lease expiry, and residual inventory. Preserve the write-once backend binding so selector changes cannot route cleanup through the wrong provider.

Escalate on ambiguous backend ownership, a revoked authority that blocks accepted cleanup, network isolation failure, or any proposed direct cloud deletion.