Applications

Databases

Every PostgreSQL database in the cluster lives in the database namespace, managed by a single CloudNative-PG operator. The same namespace hosts the Dragonfly and ClickHouse operators, so all stateful data engines are in one place.

There are three PostgreSQL clusters serving 21 databases. Applications do not get a cluster of their own; they get a tenant: a database and a login role on one of the two shared clusters.

Architecture

flowchart TD
    subgraph DB["database namespace"]
        OP[CNPG operator<br/>WATCH_NAMESPACE: database]
        BARMAN[barman-cloud plugin]
        OS[ObjectStore: garage<br/>s3://cnpg/]
        CSS[ClusterSecretStore<br/>cnpg-secrets-database]

        SHARED[("shared<br/>pg18 · 15 tenants")]
        AI[("ai<br/>vectorchord 18 · 5 tenants")]
        IMMICH[("immich<br/>vchord 16 · dedicated")]

        TS[Tenant Secrets<br/>shared-*, ai-*]
    end

    subgraph Apps["Application namespaces"]
        A1[19 apps via the<br/>cnpg-db-shared component]
        A3[immich · inline ExternalSecret]
    end

    OP -->|manages| SHARED
    OP -->|manages| AI
    OP -->|manages| IMMICH
    SHARED --> OS
    AI --> OS
    IMMICH --> OS
    BARMAN -.->|WAL archive + base backup| OS

    TS -->|username/password<br/>reconciles DatabaseRole| SHARED
    CSS -->|reads| TS
    A1 -->|ExternalSecret| CSS
    A3 -->|ExternalSecret| CSS

    classDef operator fill:#7c3aed,stroke:#5b21b6,color:#fff
    classDef cluster fill:#00b894,stroke:#00a381,color:#fff
    class OP operator
    class SHARED,AI,IMMICH cluster

Clusters

Definitions live in kubernetes/apps/pitower/database/clusters/.

Why ai is separate from shared: memini's backend runs CREATE EXTENSION IF NOT EXISTS vchord CASCADE at startup and needs the extension binaries present. vchord is a background-worker extension, so it must be in shared_preload_libraries at postmaster start, which is cluster-wide and not something to impose on tenants that do not need it.

Why immich stays dedicated: its migrations create cube, earthdistance and vchord themselves, which needs SUPERUSER. That is defensible on a single-tenant cluster and not on a shared one.

Each cluster exposes three services: <cluster>-rw (primary), <cluster>-ro (replicas) and <cluster>-r (any instance). Applications use -rw.

Cluster configuration notes

spec:
  instances: 2
  imageName: ghcr.io/cloudnative-pg/postgresql:18
  primaryUpdateMethod: switchover
  postgresql:
    parameters:
      max_connections: "200"
  resources:
    requests:
      cpu: 250m
      memory: 1Gi
  storage:
    size: 10Gi
    storageClass: openebs-hostpath

Three of those are deliberate and easy to get wrong:

• primaryUpdateMethod: switchover. On any PodSpec change CNPG defaults to restarting the primary in place. With openebs-hostpath the data is node-local, so a primary that cannot reschedule onto its own node leaves the cluster with no primary and writes down, with no automatic failover. Switchover promotes the replica first.
• requests only, no memory limit. A memory limit caps the cgroup's page cache, which for Postgres means the shared buffer working set gets evicted and reads fall through to disk.
• max_connections: "200" is a cluster-wide budget, the sum across all of its tenants, not a per-app allowance. An app with an unbounded connection pool can starve every other tenant; see Adding a database.

Tenancy model

A tenant is two CRs plus a Secret, in one file under kubernetes/apps/pitower/database/tenants/:

apiVersion: external-secrets.io/v1
kind: ExternalSecret
metadata:
  name: shared-miniflux
spec:
  secretStoreRef:
    kind: ClusterSecretStore
    name: infisical
  target:
    name: shared-miniflux
    template:
      type: kubernetes.io/basic-auth      # CNPG requires exactly this type
      metadata:
        labels:
          cnpg.io/reload: "true"          # pick up password changes immediately
      data:
        username: miniflux                # must match the role name
        password: "{{ .password }}"
        host: shared-rw.database.svc.cluster.local
        port: "5432"
        dbname: miniflux
  data:
    - secretKey: password
      remoteRef:
        key: /database/tenants/MINIFLUX_PASSWORD
---
apiVersion: postgresql.cnpg.io/v1
kind: DatabaseRole
metadata:
  name: miniflux
spec:
  cluster:
    name: shared
  name: miniflux
  login: true
  databaseRoleReclaimPolicy: retain       # deleting this CR must not DROP ROLE
  passwordSecret:
    name: shared-miniflux
---
apiVersion: postgresql.cnpg.io/v1
kind: Database
metadata:
  name: shared-miniflux
spec:
  cluster:
    name: shared
  name: miniflux
  owner: miniflux
  databaseReclaimPolicy: retain           # deleting this CR must not DROP DATABASE

Three rules hold this together:

The tenant Secret does double duty. CNPG reads username/password from it to reconcile the DatabaseRole, and the consuming application reads all five keys out of the same object. Because it carries host/port/dbname too, the consumer component needs no namespace and no service FQDN of its own: moving a cluster or renaming a service changes the source Secret, not every consumer.

Note the key is username, not the user that CNPG's own <cluster>-app Secrets use.

Passwords live in Infisical at /database/tenants/<APP>_PASSWORD.

Consuming a database

Nineteen apps use the cnpg-db-shared kustomize component. It reads the tenant Secret through the cnpg-secrets-database ClusterSecretStore and emits a single Secret in the app's namespace:

Wiring it up is a component reference plus four literals:

components:
  - ../../../../components/cnpg-db-shared
configMapGenerator:
  - name: cnpg-db-config
    literals:
      - APP_NAME=miniflux
      - SECRET_NAME=miniflux-db-secret
      - CNPG_SECRET_KEY=shared-miniflux
      - DB_SCHEME=postgresql

Tenant map

The exception

Not every consumer uses the component:

• immich reads the CNPG-generated immich-app Secret with an inline ExternalSecret against the same cnpg-secrets-database store, because it is the one cluster that is not multi-tenant and so has no hand-authored tenant Secret to read.

Adding a database

The whole recipe, for an app that needs a new database on shared.

1. Store a password

pw=$(LC_ALL=C tr -dc 'A-Za-z0-9' </dev/urandom | head -c 64)
env -u INFISICAL_TOKEN infisical secrets set "MYAPP_PASSWORD=$pw" \
  --path=/database/tenants --env=prod --silent

2. Add the tenant file

Copy tenants/miniflux.yaml, change the five occurrences of the app name and the Infisical key, and add it to tenants/kustomization.yaml.

Pick the cluster: ai if it needs vector support, shared otherwise.

3. Wait for the CRs to apply

kubectl --context=admin@pitower -n database get database,databaserole

Both must report APPLIED: true before the app starts.

4. Point the app at it

Add the component and the four literals shown above, then consume DB_URL (or the individual keys) from <app>-db-secret.

Add reloader.stakater.com/auto: "true" to the controller so a password change restarts it.

5. Cap the connection pool if the app needs it

max_connections: "200" is shared across every tenant. If the app's pool is unbounded by default (Forgejo's is), set a limit as part of onboarding:

MAX_OPEN_CONNS: 20
MAX_IDLE_CONNS: 5

6. Backups need nothing

The tenant inherits the cluster's WAL archiving and nightly base backup. There is no per-tenant backup configuration.

Backups

A single ObjectStore named garage serves every cluster, via the barman-cloud CNPG-I plugin (the in-tree backup.barmanObjectStore field is deprecated as of CNPG 1.26).

apiVersion: barmancloud.cnpg.io/v1
kind: ObjectStore
metadata:
  name: garage
spec:
  retentionPolicy: 30d
  configuration:
    destinationPath: s3://cnpg/
    endpointURL: https://s3.wibrow.dev
    wal:
      compression: gzip
    data:
      compression: gzip

Barman namespaces each cluster's backups under its own serverName, which defaults to the cluster name, so shared and ai land in s3://cnpg/shared/ and s3://cnpg/ai/ without extra configuration.

Each Cluster opts in with a plugin entry:

plugins:
  - name: barman-cloud.cloudnative-pg.io
    isWALArchiver: true
    parameters:
      barmanObjectName: garage

WAL archiving is continuous and independent of the nightly ScheduledBackup; the schedule only sets how far back a PITR has to replay from.

Backups run target: prefer-standby so they do not compete with tenant traffic on the primary. Storage is Garage (system/garage); the cnpg key is scoped to the cnpg bucket.

kubectl --context=admin@pitower -n database get backup
kubectl --context=admin@pitower -n database get scheduledbackup

Storage and placement

All clusters use openebs-hostpath, backed by local NVMe. Because that data is node-local, a kustomize patch pins every Cluster to worker-05/worker-06 with required pod anti-affinity so the two instances never share a node:

spec:
  affinity:
    podAntiAffinityType: required
    topologyKey: kubernetes.io/hostname
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
          - matchExpressions:
              - key: kubernetes.io/hostname
                operator: In
                values: [worker-05, worker-06]

Adding a third database node means editing this patch. The patch is applied to every Cluster in the overlay by kind, so it cannot be forgotten for a new one.

Monitoring

Every Cluster sets monitoring.enablePodMonitor: true, so Prometheus scrapes each instance directly. The operator itself is scraped via monitoring.podMonitorEnabled: true in its Helm values.

The operator chart's bundled Grafana dashboard is disabled (grafanaDashboard.create: false); dashboards are vendored centrally instead; see Grafana.

Ad-hoc access

database-toolbox (Google genai-toolbox, in the ai namespace) fronts every database in the namespace over MCP. Its init container enumerates the credential Secrets in database and generates one source per database, so a new tenant is picked up on the next pod restart with no configuration change.

It connects as each database's owning role, so its execute_sql tool can write. Only the *_list_tables half is exposed through garrison's allowlist.

Direct psql access, for when that is not enough:

kubectl --context=admin@pitower -n database exec -it shared-1 -c postgres -- \
  psql -U postgres -d miniflux

Other engines

The database namespace also hosts two non-Postgres operators:

Dragonfly instances are declared as Dragonfly CRs next to the app that uses them, not here.

See also

• CNPG Consolidation: how this layout came to be
• Backup & Restore
• Garage S3
• External Secrets
• OpenEBS