GCP preview-env loop (mldp7 base + pr-999) — 2026-09-14/15

Live rehearsal of the preview-environment lifecycle on the certified GCP origin (GKE Autopilot, europe-west1): long-running base up, PR preview created from it, health-verified, guards exercised, both torn down.

Scope

• Base env mldp7-preview: full acceptance matrix (13 cells).
• Preview env mldp7-pr-999: bootstrap, deploy, health, guard checks, destroy. Preview identity derives from repository plus PR number.
• Teardown completed 2026-09-15; all billable resources deleted.

Loop rows

Step	Outcome
Base mldp7-preview acceptance	13/13 PASS (bootstrap:wif, composer:sm-write, composer:sm-read, day2:secrets, day2:state, day2:logs, day2:exec, day2:health, search:health, migrate:dump, deploy:candidate, deploy:repeat, cost:estimate)
Preview pr-999 bootstrap + deploy	Deployed; Magento base URLs corrected post-deploy, runtime health confirmed
Production/protected destroy guard	Refusal verified: protected env destroy without --yes does not proceed
env sweep --dry-run	Reports the expired preview without touching the long-running base
Preview destroy	Compute, data, buckets, service accounts, subnets deleted; VPC peering delete refused by GCP (see below)
Base destroy	Cluster, Cloud SQL, Valkey, buckets, service accounts, routers, subnets deleted

Teardown notes

The Pulumi passphrase file for this rehearsal lived in /tmp and was reaped before teardown, so magelift destroy could not unlock the stacks and the teardown ran as ordered gcloud deletes (cluster, SQL, Valkey, buckets, service accounts, NATs, routers, subnets). All billable resources are gone; verification sweeps found no clusters, SQL instances, Valkey instances, buckets, service accounts, load-balancer parts, DNS zones, or firewall rules for either env.

Two Service Networking peerings (mldp7-preview-net, mldp7-pr-999-net) plus their empty VPCs and reserved ranges remain: GCP refuses the peering delete with "Producer services ... are still using this connection" although no consumers exist (no SQL, no Memorystore/Valkey, no Redis in any region). This is GCP-side stuck state, not a MageLift orphan: peerings, empty VPCs, and unused ranges cost nothing. Retried across 14+ hours, most recently at intent close; still refused. Future sessions may re-attempt the delete.

Spend

Both envs lived under one day inside the acceptance project; no separate budget line was tracked for the rehearsal.

Sources

• Base transcript: gcp-loop-mldp7.log (13 cell-done ... PASS rows).
• Preview deploy log: mldp7-pr999-deploy.log.
• Session transcript: preview CLI steps plus guard checks plus teardown.