GCP foundation ownership and operations
GCP foundation ownership and operations
Decision
The development foundation was applied with owner authorization on 2026-10-12. Terraform state was migrated to its private GCS backend. Build federation is enabled for the merged owner-triggered, main-only image workflow; GitHub still has no Terraform apply permission. Application/database deployment and live OIDC exchange remain unverified. See the evidence ledger.
Bootstrap instructions own the exact commands. The delivery contract owns the selected architecture.
Applied bootstrap identity and state boundary
| Identity | Granted in this slice | Later owner |
|---|---|---|
| Owner bootstrap | Existing separately supplied administrative authority; not minted here | Owner applies foundation and privately reviews state/plans |
| Builder | Writer on only the development image repository; narrowly scoped WIF impersonation, enabled for development | #27 builds/scans/promotes digests; no state, runtime impersonation or secrets |
| Deployer | Object admin on only application-state bucket; no WIF or cloud deployment rights | #27 adds minimum resource roles and a reviewed promotion gate |
| Migrator | Account only | #26 owns DDL role and bounded migration invocation |
| API, frontend | Separate accounts only | #26/#27 own database/connectivity and ingress privileges |
| Publisher, consumer | Separate accounts only | #7 owns task enqueue/handling and SQL receipt rights |
| Scheduler, task-invoker | Separate accounts only | #7 scopes invoke/token permissions to the specific handlers |
The table records the initial applied bootstrap. New, unapplied #27 source moves deployer state access to a third runtime bucket and adds separately disabled deployment federation; the rollout contract owns its permissions and promotion gate. Foundation state has no application-deployer grant. Both buckets enforce uniform access, public-access prevention, versioning and Terraform deletion guards. These grants do not erase inherited owner/org permissions; verify effective IAM before apply. Infisical impersonation/secret containers and values stay in #11; this foundation creates no unused secret or machine credential.
Promotion and recovery
Private main protection returned HTTP 403 on 2026-10-11; environments list was empty. Neither a workflow label nor a manual convention is an enforced approval. Before #27 enables publishing, owner reviews passing checks and initiates the exact main workflow. Provider checks actor, immutable repository/owner IDs, event, ref and workflow path. This limits build authority; repository administrators remain trusted. No automatic apply or deployment follows a push, merge or issue label.
Only Terraform may own configured runtime revisions. #27 must serialize each dev migration/deployment, accept an already scanned digest, stop on migration/readiness/ auth failure and roll back only to a schema-compatible digest. Runtime DDL is denied; failed migrations never trigger force/automatic down. Retain last known good tagged images; cleanup deletes only untagged images older than seven days.
On federation incident, owner disables pool/provider before reviewing trust; already-issued access tokens can remain valid until expiry. Revoke scoped grants where immediate containment is required. For state loss, owner restores a version privately and verifies it against resources; never delete buckets or force-unlock without identifying the lock owner. Default tests cannot exercise these cloud paths.
Capacity, costs and acceptance inputs
Initial proposed DB pool budget for #26: API 5 per instance, publisher 2, consumer 2; maxima 2/1/1 give 14 connections. Allow a second revision of each (28), two migrator and 5 administrative connections: 35 total (migration lock plus execution). Verify the actual server connection limit and load behavior; frontend has no SQL pool. This is a proposal for #26, not an implemented pool setting or permission to exhaust the server.
Development recovery objectives proposed for owner acceptance: restore within four hours and lose at most 24 hours of synthetic work through the backup path; PITR is additionally tested. One-minute publisher activation is a schedule, not a delivery SLA; measure enqueue/completion latency and investigate oldest pending work above five minutes. No live-bookkeeping recovery objective is inferred from these defaults.
Budget alerts at 50%, 80% and 100% of USD 50 exclude credits and apply to the selected project. They notify default billing recipients; owner must verify those recipients and delivery in the cloud proof. Alerts never enforce a spend cap.
Reviewed Iowa SQL rates give an illustrative 730-hour baseline: db-g1-small at $0.035/hour = $25.55; 10 GiB SSD at $0.000232877/GiB-hour = about $1.70; 10 GiB of used backup storage at $0.000109589/GiB-hour = about $0.80. That subtotal is $28.05 before PITR growth, restore resources and other services. Shared-core instances have no Cloud SQL SLA. See G10 in the source register.
Planning envelope (not a quote): reserve $30 for database compute/initial storage, $5 for backups/PITR and restore testing, $5 for application/publisher/consumer compute, $1 for scheduler/tasks, $3 for state/artifacts/logging and $3 for authentication/secrets/ networking, leaving $3 contingency. These are allocation limits, not verified SKU estimates. Require a regional calculator estimate before apply; do not assume free allowances are unused across this billing account. Bank subscriptions are excluded. Observed seven-day spend, synthetic restore timing, backlog latency and the 1,000-post proof belong to #12. Project and billing metadata are now established and #5 is closed in GitHub. Application auth, actual trust-denial tests, cost measurements and recovery acceptance remain separate cloud evidence gates.
Acquisition bootstrap state
The hosted acquisition slice adds a fourth private versioned state bucket,
PROJECT-jayburd-dev-workers-state, for the owner-operated workers root. The
application deployer's existing runtime-state grant does not cover it. Apply the
reviewed foundation addition before installing that backend; a prefix in the shared
runtime bucket is not a substitute for IAM separation. Mocked tests verify the
configuration; this change has not provisioned the bucket. See the
hosted platform boundary.