approval-engine/deploy/README.md
tegwick be1a388a84 Enforce declared human controls at approval binding
Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a07ff8-19d0-7820-b4d0-1353833cb7fc
2026-09-10 19:26:12 +02:00

50 lines
2.8 KiB
Markdown

# Deployment gates
The checked-in StatefulSet is a reviewed release input, not evidence of a live
deployment. SQLite is intentionally limited to one replica, `ReadWriteOnce`
storage, and an `OnDelete` update: never start two writers against a copied
database.
**Image: pinned, published, not rolled out.** Both `image:` references already
carry the immutable release digest
`sha256:73333f5ceb55e48192e3095cb2e2a741cdc6ff0be2f18128301072b4a6b6eb9d`
(registry tag `0.1.0`, OCI index; linux/amd64 manifest `sha256:2064d537…`),
pinned at `b51d174`. There is no placeholder left to replace. Both references
MUST stay identical and MUST stay digests — a tag here would let the init
container and the server run different code against one database.
**The pinned image predates schema v5.** The published artifact recorded in
[`docs/image-scan-2026-09-06.md`](../docs/image-scan-2026-09-06.md) carries
`LATEST_SCHEMA_VERSION = 3`; this repository is now at 5
(`entries.principal_type` plus `approvals.human_control`, see [storage-operations.md](../docs/storage-operations.md)).
The pinned pair is self-consistent — that image migrates to 3 and serves 3 — so
nothing is broken by leaving it pinned, but a rollout that must enforce declared human controls or carry approver
principal-type evidence requires cutting a new image at step 3 below. The
`migrate` init container then performs the additive upgrade on the existing
volume; `tests/test_deploy_manifest.py` holds this acknowledgement so the drift
cannot go quiet.
Gates 1 and 2 below are the outstanding ones; nothing is deployed today.
Before applying:
1. Register the exact `approval-engine` audience, caller scopes, and service
clients in KeyCape. Confirm the in-cluster JWKS endpoint and configured
issuer match the manifest.
2. Register `approval-engine` as an audit-core sender, add matching audit-core
ingress, and provision `approval-engine-audit` through the credential owner.
The Secret must contain key `audit-token`; never commit its value.
3. For a new release only: `make image-release` (build, scan, push — the scan
gate fails the build on any CRITICAL or HIGH, so a failing scan blocks the
push), then pin both `image:` references to the returned registry digest and
run `make deploy-dry-run`. The current digest is already pinned; skip to 4
unless you are cutting a new image.
4. Take a verified backup. Roll out by deleting the sole pod, then prove
readiness, restart persistence, outbox drain, heartbeat, and restore.
5. Enable a caller namespace only by applying label
`railiance.io/approval-engine-client=true`; JWT scope checks remain the inner
boundary.
The audit-core receiver-side registration, cadence findings, and accepted-count
reconciliation are tracked by `AUDIT-WP-0009-T04/T06/T09` and cannot be closed
from this repository.