railiance-platform/docs/postgresql-ha.md

PostgreSQL HA — retired

Status: retired 2026-08-15 (RAILIANCE-WP-0016 item 14)
Superseded by: CloudNativePG clusters in databases (apps-pg, forgejo-db, platform-pg, and the other CNPG instances).
Do not deploy. make pg-deploy is fail-closed.

The Bitnami postgresql-ha (repmgr + pgpool) stack never ran as a live S3 service on railiance01. There is no platform namespace and no matching pods. Gitea’s historical subchart in S2 was replaced by forgejo-db. Keep this file as the retirement record and the pgpool-password incident note. New databases go through docs/apps-pg.md or rapp-postgres.

---

Architecture

Apps (S5)
  └── pgpool (load balancer / connection pooler)
        ├── postgresql-0  [Primary  — repmgr]
        ├── postgresql-1  [Standby  — repmgr]
        └── postgresql-2  [Standby  — repmgr]

• pgpool-II distributes reads across standbys, routes writes to primary
• repmgr handles automatic failover if the primary disappears
• All pods in platform namespace; app pods connect via pgpool service

Connection string pattern

postgresql://DBUSER:DBPASS@postgresql-ha-pgpool.platform.svc.cluster.local:5432/DBNAME

Replace DBUSER, DBPASS, DBNAME with the database-specific credentials.

---

Initial deployment

Prerequisites

• railiance-cluster converged (make smoke passes)
• SOPS age key accessible: sops -d helm/postgresql-ha-values.sops.yaml returns plaintext
• helm repo add bitnami https://charts.bitnami.com/bitnami && helm repo update done on the node

Steps

# 1. Ensure the platform namespace exists
kubectl create namespace platform --dry-run=client -o yaml | kubectl apply -f -

# 2. Deploy (from railiance-platform/)
make pg-deploy

# 3. Verify
make pg-status
# Expected: 3 postgresql pods + 1 pgpool pod, all Running

# 4. Smoke test
make smoke

---

Creating a new database for an app

# Connect via pgpool
kubectl exec -it -n platform \
  $(kubectl get pod -n platform -l app.kubernetes.io/component=pgpool -o name | head -1) \
  -- psql -U postgres

# Inside psql:
CREATE DATABASE myapp;
CREATE USER myapp WITH PASSWORD 'strong-password';
GRANT ALL PRIVILEGES ON DATABASE myapp TO myapp;
\c myapp
GRANT ALL ON SCHEMA public TO myapp;
\q

Add the user password to the app's own secrets (managed in the app's repo, not here). The connection string for the app will be:

postgresql://myapp:strong-password@postgresql-ha-pgpool.platform.svc.cluster.local:5432/myapp

---

Password rotation

1. Update the password in the plaintext values template
2. Re-encrypt: sops -e -i helm/postgresql-ha-values.sops.yaml
3. Upgrade: make pg-deploy
4. Update the app's connection secret to match
5. Rolling restart the app pods to pick up the new connection

---

pgpool-password — critical note

The postgresql.pgpoolPassword value in the Helm chart maps to the pgpool-password key in the postgresql-ha-postgresql Kubernetes Secret. The pgpool container mounts this key at startup; if it is absent, pgpool enters CrashLoopBackOff with no log output.

This was the root cause of the 2026-03-10 incident (RAIL-BS-WP-0003).

Always verify after helm upgrade:

kubectl get secret -n platform postgresql-ha-postgresql \
  -o jsonpath='{.data.pgpool-password}' | base64 -d && echo
# Must print a non-empty string

---

HA failover test

Per Decision D3, any change to this service requires a passing failover test:

# From railiance-cluster/
make test-ha-failover GITEA_URL=https://<gitea-hostname>

The test kills the primary PostgreSQL pod and asserts:

1. repmgr promotes a standby within 60s
2. All pods return to Running within 120s
3. pgpool returns to Running (catches the missing-key bug)

---

Backup

Platform backup (PostgreSQL logical dump) is handled by the railiance-backup tool in this repo:

make backup

This produces an age-encrypted dump uploaded to Nextcloud. For cluster-level backup (etcd, kubeconfig), see railiance-cluster/.

---

Scaling to 3 nodes (ThreePhoenix)

When Railiance02 and Railiance03 join the cluster:

1. Switch StorageClass from local-path to longhorn in the values file
2. Change postgresql.podAntiAffinityPreset from soft to hard
3. Run make pg-deploy — Helm rolling update spreads pods across nodes
4. Run make test-ha-failover to confirm HA is genuine (not just replicated on one node)