# apps-pg resource evidence

Work record: `RAILIANCE-WP-0016`
Resource: `resource:railiance:apps-pg`
Captured: `2026-08-14T00:04:19Z` (refresh:
`scripts/capture-apps-pg-evidence.py -o docs/evidence/RAILIANCE-WP-0016-apps-pg-observation.json`)
Machine-readable companion: `docs/evidence/RAILIANCE-WP-0016-apps-pg-observation.json`

This is the platform-owned evidence `RESOURCE-WP-0003-T04` asked for.
It does not contain credentials, DSNs, or application rows. It is not a
fin-hub ledger: no EUR amounts are invented here. resource-control can
forecast and allocate from these technical facts; booked spend stays in
fin-hub.

## T01 — Capacity and utilization

| Fact | Value | Kind | Source |
| --- | --- | --- | --- |
| Instances | 1 provisioned, 1 ready | declared + observed | Cluster `apps-pg` |
| Image | `ghcr.io/cloudnative-pg/postgresql:16` (live 16.13) | declared | Cluster spec / `version()` |
| Created | 2026-08-09T00:15:53Z | observed | Cluster metadata |
| Phase | Cluster in healthy state | observed | Cluster status |
| CPU / memory requests | unset (BestEffort) | declared | empty `spec.resources` |
| CPU / memory observed | 9m / 102Mi | observed | `kubectl top` |
| Storage provisioned | 10Gi, `local-path`, RWO | declared | PVC `apps-pg-1` |
| PGDATA used | 619 083 120 B (~590.4 MiB) | observed | `du -sb …/pgdata` |
| WAL since init | 587 202 560 B (~560 MiB) | observed | `pg_wal_lsn_diff` |
| max_connections | 100 | declared | `SHOW max_connections` |
| Sessions at capture | 7 (1 `postgres` + 6 background) | observed | `pg_stat_activity` |

Consumer catalogs, five days after commission:

| Database | Workload | Bytes | xact_commit | backends |
| --- | --- | --- | --- | --- |
| `coulomb_social_db` | coulomb-social | 8 952 855 | 14 754 | 0 |
| `vergabe_db` | vergabe-teilnahme | 7 668 239 | 0 | 0 |
| `apps_meta` | platform bootstrap | 7 830 551 | 14 374 | 0 |
| `postgres` | engine catalog | 7 830 551 | 146 383 | 1 |

`vergabe_db` has never been used. `coulomb_social_db` is the only consumer
with writes (982 inserts). Cache hit on that database is >99.9%. There is
no I/O history older than cluster creation and no metrics retention;
saturation proxies are the snapshot above plus WAL growth.

Headroom: ~9.4 GiB of the 10Gi PVC is free. The instance has no CPU or
memory request, so it does not reserve cluster capacity. Observed use is
idle-class.

## T02 — Recovery and operations

| Check | Result |
| --- | --- |
| Cluster Ready | true since 2026-08-09T00:16:16Z |
| ContinuousArchiving | true — CNPG `wal-archive` to the instance manager, **on-node** |
| CNPG `Backup` / `ScheduledBackup` | none |
| Option A logical dump | **not a target**. Default list covers CoulombCore `apps-pg` and railiance01 `forgejo-db` / `net-kingdom-pg` / `state-hub-db` / `user-engine-pg` only |
| Restore drill for this cluster | none recorded |
| Replica / pooler | not configured (single instance) |
| Failure domain | `railiance01` + `local-path`. Host loss is data loss |

Recovery objectives: **RPO unbounded, RTO undefined**. This is unresolved
local-storage risk, not verified recovery. `docs/apps-pg.md` already said
backup was follow-up work; this observation confirms it has not landed for
the railiance01 cluster.

Maintenance: image tag `postgresql:16` is a major pin, not a patch pin.
The live minor is 16.13. No upgrade or incident record exists for this
cluster. Cadence is unscheduled.

Labor (high uncertainty, not a timesheet):

| Class | Estimate | Basis |
| --- | --- | --- |
| Setup | 6 operator-hours | RAILIANCE-WP-0003 cluster + two consumer onboardings + docs, 2026-08-09 |
| Recurring | 0.5 operator-hours / month | ad-hoc `make apps-pg-status`; no backup, restore, or upgrade job |
| Incidents | 0 recorded | none in workplans or Cluster conditions |

resource-control may convert hours at its own labor rate. Do not treat
these hours as booked cost.

## T03 — Allocation drivers

`apps-pg` is a shared platform service. The recommended first method is
the one resource-control already sketched, with the residual made
explicit.

**Method `apps-pg-dbbytes-v1`**

- Driver: `database_gb` = `pg_database_size` of each **declared consumer**
  database.
- Denominator: sum of those consumer sizes (16 621 094 B at capture).
- Shares at capture: coulomb-social **0.5386**, vergabe-teilnahme **0.4614**.
- Residual: PGDATA minus all non-template databases = 586 800 924 B
  (**94.8%** of used volume). Attribute to
  `platform:apps-pg-overhead`. This is WAL plus engine catalogs, not
  missing consumer data.
- Uncertainty: **high**. Five days of life, empty `vergabe_db`, no history.

Candidates considered and not selected as v1:

| Driver | Why not first |
| --- | --- |
| Connections | All consumers idle at capture; peak is our own `psql` |
| `xact_commit` / tuples | Would assign ~100% to coulomb-social and 0% to a reserved empty database |
| Equal instance share | Fair while data is tiny, but it hides growth; keep as fallback `apps-pg-instance-v1` if residual stays above 50% after the first month |

Shared overhead stays unattributed to a consumer on purpose. Converting
the residual into booked EUR is fin-hub / resource-control work. Adding a
third consumer is a new row in `consumers` plus a Database CR; the method
does not change.

## What this is not

- Not a cost figure. The host bill belongs to `resource:hosteurope:railiance01`.
- Not a backup program. Covering railiance01 `apps-pg` is a later platform
  change, gated on the same object-store decision as `platform-pg`.
- Not application telemetry. No table counts, no tenant identifiers, no
  Secret values.