ADR-0015: NetKingdom Railiance packaging and relational platform
All checks were successful
CI Smoke / host-smoke (push) Successful in 0s
CI Smoke / container-smoke (push) Successful in 1s

Settles two questions raised by bringing NetKingdom under Railiance
governance:

1. Separate rapp-* repos per engine (rapp-tenant-engine, rapp-user-engine),
   following repository-axes.md's one-workload rule. The decisive property is
   independent rollback -- a single rapp would need one rollback contract
   across independently versioned services. secrets-engine is not packaged as
   a rapp: it has no deployed workload.

2. CloudNative PG via rapp-postgres is the default relational platform for
   production. Per-workload SQLite-on-a-PVC is dev/test only, and
   rail-kubernetes wave-1 does not support the persistent-storage contract it
   depends on. tenant-engine migrates; its TenantStore Protocol makes this a
   backend swap behind an existing seam.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
tegwick 2026-08-11 14:46:33 +02:00
parent 4511a43f7a
commit 1bb7bce99b

View file

@ -0,0 +1,166 @@
# ADR-0015 - NetKingdom Railiance Workload Packaging and Relational Platform
**Status:** Accepted
**Date:** 2026-08-11
**Deciders:** Bernd Worsch, Claude
## Context
NetKingdom's runtime services are deployed today outside the Railiance
reef/rail/`rapp` model. `tenant-engine` and `user-engine` run on the coulomb
substrate with digest-pinned images
(`forgejo.coulomb.social/coulomb/{tenant,user}-engine@sha256:…`), but their
Kubernetes manifests live in this repo at `sso-mfa/k8s/<service>/runtime.yaml`
— a canon repo holding runtime YAML — applied imperatively, with
`verify-t0*.sh` scripts as verification. Neither service declares
`railiance/app.toml`, appears in `reef-railiance/bindings/rapps.yaml`, or is
reconciled by a GitOps controller.
The decision to bring NetKingdom under Railiance governance forces two
questions this ADR settles.
**Packaging granularity.** `railiance-master/docs/repository-axes.md` maps
`rapp-*` to "one Railiance-managed workload" and states that a `rapp` "does
not replace the responsibility repo that owns the broader domain. It is
explicitly about managed wrapping, not ownership." Precedent exists in two
shapes: `rapp-qonto` (`ownership_repo: qonto-assistant`, one workload, thin
package repo pointing back at the domain repo) and `rapp-postgres`
(`ownership_repo: railiance-platform`, one workload with a `consumers:` list
and a `consumer_contract`). The open question was whether NetKingdom's engines
are one workload or several.
**Relational storage.** The engines have diverged. `user-engine` already runs
CloudNative PG (`kind: Cluster` in its own namespace) with NetworkPolicies and
a `migrations/` surface. `tenant-engine` runs SQLite on a 1 Gi RWO PVC, chosen
in TEN-WP-0004 for expedience. `railiance-platform` (S3) declares CloudNative
PG the canonical database operator, and `rapp-postgres` packages it with a
`database-per-consumer` boundary unit and OpenBao dynamic credentials.
Separately, `rail-kubernetes`'s wave-1 contract explicitly states the rail must
**not** assume "a generic persistent-storage contract" — so a per-workload PVC
is unsupported by the rail regardless of which database is chosen.
TEN-WP-0005 was drafted against PostgreSQL and implemented against SQLite,
because T05 must roll out against the runtime that exists. That workplan's
`TenantStore` Protocol was kept as an explicit seam for this decision.
## Decision
### 1. Separate `rapp-*` repos per engine
Each deployed NetKingdom engine gets its own managed workload package repo:
```text
rapp-tenant-engine ownership_repo: tenant-engine
rapp-user-engine ownership_repo: user-engine
```
Following `rapp-qonto`'s shape: the `rapp` owns Railiance packaging, rail
compatibility, workload-specific smoke/health checks, rollout and rollback
expectations, secret references, and dependency declarations. The engine repo
retains domain ownership; `net-kingdom` retains canon.
This follows the one-workload rule rather than working around it. The engines
are already independent on every axis a `rapp` declaration must state:
separate namespaces, separate image digests, separate release cadence, and
(until decision 2 lands) different storage. A single `rapp-netkingdom` would
have to declare one `workload_identity`, one `rollout_contract`, and one
`rollback_contract` across services that genuinely differ.
The decisive property is **independent rollback**: rolling back a bad
`tenant-engine` revision must not force a `user-engine` revision change. A
single rapp would make that either impossible or fictional.
Rollout ordering between the engines is a declared constraint between rapps,
not a reason to merge them: `tenant-engine` before `user-engine`, following the
dependency direction.
### 2. `secrets-engine` is not packaged as a `rapp`
`secrets-engine` has no deployed workload — no Containerfile, no Makefile, no
Kubernetes manifests. It is a catalog/policy/workflow layer over OpenBao, whose
packaging already belongs to `rapp-openbao` with custody and lane policy in
`railiance-platform`.
A `rapp` wraps a workload; `secrets-engine` has none to wrap. Revisit only if
it becomes a runtime service, at which point this decision reopens for that
repo alone and not for the packaging model.
### 3. CloudNative PG is the default relational platform for production
Production NetKingdom services requiring relational storage use CloudNative PG
via `rapp-postgres`, consuming the `database-per-consumer` boundary unit and
the `openbao-dynamic-database-credential` lane, with tenant-keying per
`business-app-service-contract_v0.1` section 1.3.
Per-workload SQLite-on-a-PVC is not a production pattern. It remains acceptable
for local development and tests.
`tenant-engine` migrates from SQLite to cnpg. Its `TenantStore` Protocol makes
this a backend swap behind an existing seam rather than a rewrite; the
lifecycle semantics proven in TEN-WP-0005 (atomic compare-and-swap, durable
idempotency receipts, forward-only migration) are the conformance bar the
PostgreSQL backend must meet, and its store-conformance suite is already
parametrised across backends to enforce exactly that.
## Consequences
- Two new `rapp-*` repos to create, each requiring a `declarations/rapp.yaml`,
a binding in `reef-railiance/bindings/rapps.yaml`, and its own readiness
evidence progression (`declared``installed``verified`
`production-approved`).
- Per-binding evidence multiplies, but per-reef evidence does not — the
`reef-production-readiness-contract` puts substrate, ingress, storage,
network, and backup evidence on the reef, once. What multiplies is the
critical-workload gate (threat model, negative authorization tests, rollback
rules, residual-risk owner), which identity-plane services warrant
individually. A shared NetKingdom threat model may be referenced by both
bindings rather than duplicated.
- Runtime manifests move out of `net-kingdom/sso-mfa/k8s/` into the respective
`rapp` repos. This repo stops holding runtime YAML and returns to canon,
standards, and ADRs.
- `tenant-engine` requires a data migration from SQLite to cnpg, including
migration of existing tenants, grants, plan assignments, and idempotency
receipts. TEN-WP-0005-T05's rollout plan is affected: it currently targets
the SQLite runtime.
- NetKingdom services become subject to the `railiance/app.toml` staged
promotion contract (Stage 1 local → Stage 2 canary → Stage 3 production),
including declared rollback commands and health endpoints. Digest-pinned
images already satisfy `digest_policy = "required"`.
- The reef must accept or mitigate its single-server and shared-control-plane
risk for each NetKingdom binding, as it must for `rapp-qonto`. The reef name
and a `high` criticality label are not evidence.
## Alternatives Considered
**One `rapp-netkingdom` for all associated repos.** Rejected. It would reduce
binding-evidence count and match the current coupled deployment (both engines'
manifests share one tree; `flex-auth` runs per-consumer instances
`flex-auth-tenant-engine` and `flex-auth-user-engine`). But it contradicts the
one-workload rule in `repository-axes.md`, and would force a single rollback
contract across independently versioned services. A rollback contract that
depends on which service failed is not a contract. This alternative would be
correct only if the engines were always promoted and rolled back as one atomic
cutover — which their separate digests and release cadences contradict.
**Keep SQLite for `tenant-engine`, standardise later.** Rejected. It leaves the
two engines operationally dissimilar, which is the opposite of the intent, and
`rail-kubernetes` does not support the per-workload persistent-storage contract
the PVC depends on. Deferring also grows the migration: every tenant, grant,
and receipt written between now and the cutover is data to move.
**A `rapp` per NetKingdom concern with a `consumers:` list, mirroring
`rapp-postgres`.** Rejected as a category error. `rapp-postgres`'s consumers
are consumers *of one workload* (PostgreSQL); NetKingdom's engines are distinct
workloads, not consumers of a shared one.
## Follow-Up
- Create `rapp-tenant-engine` and `rapp-user-engine`; move runtime manifests
out of `net-kingdom/sso-mfa/k8s/`.
- Add both bindings to `reef-railiance/bindings/rapps.yaml` at `declared`.
- Add `railiance/app.toml` to `tenant-engine` and `user-engine`.
- Open a `tenant-engine` workplan for the cnpg backend and data migration;
reconcile with TEN-WP-0005-T05, whose rollout currently targets SQLite.
- Confirm whether `secrets-engine` is intended to remain a non-deployed control
layer. This ADR assumes it is.
- Record the NetKingdom-wide threat model that both bindings will reference.