feat(db): switch to discourse/postgres image with install-user + checksum adapter

Replace the bitnami-era pgvector:pg17 db + hand-rolled pg_upgrade entrypoint
with discourse/postgres:pg18 (pgvector + discourse's auto-upgrade layer, as
suggested on coop-cloud/discourse#16). The image does the heavy lifting
(installs old binaries, runs pg_upgrade into the versioned PGDATA); a thin
cc-db-entrypoint.sh wrapper fills the two gaps it leaves:

- secrets: inject DB_PASSWORD/POSTGRES_PASSWORD from the docker secret (the
  image reads them from env, no *_FILE support);
- install user: detect the old cluster's bootstrap superuser (oid 10) and
  export POSTGRES_USER so pg_upgrade + the new cluster's initdb match it. Real
  deployments differ (bitnami-origin clusters install as 'postgres' + a
  'discourse' app role; others as 'discourse'). The image hardcodes
  --username=$POSTGRES_USER and never detects this, so the adapter is required;
- checksums: pg18's initdb enables data checksums by default but pg13-17
  clusters here have them off, and pg_upgrade requires a match -> initdb the new
  cluster with --no-data-checksums unless the old one reports them on.

Other changes:
- mount postgresql_data at /var/lib/postgresql (versioned PGDATA .../18/docker)
- pg_backup.sh: detect the superuser at runtime; fix paths for the new layout
- bump DB_ENTRYPOINT_VERSION v6, PG_BACKUP_VERSION v3 (immutable swarm configs)
- drop entrypoint.postgres.sh.tmpl

Verified on cctest: upgrade from an existing pg17 cluster (install user
'postgres') -> pg18, all data preserved, serves over HTTPS via Traefik.

.env.sample

@@ -21,9 +21,3 @@ DISCOURSE_DEVELOPER_EMAILS=admin@example.com
 #SECRET_SMTP_PASSWORD_VERSION=v1
 
 SECRET_DB_PASSWORD_VERSION=v1
-
-# Postgres bootstrap superuser (the cluster's "install user"). Defaults to
-# `postgres`, which matches fresh installs and bitnami-origin clusters. Only set
-# this if you are upgrading a cluster that was bootstrapped with a different
-# superuser (e.g. `discourse`) — a postgres major upgrade fails unless it matches.
-#POSTGRES_USER=postgres

README.md

@@ -47,15 +47,10 @@ abra app run YOURAPPDOMAIN app discourse admin create
 
 Handled automatically by the [`discourse/postgres`] image (pgvector + an
 auto-upgrade layer). On deploy it finds an older cluster, installs the old
-binaries and runs `pg_upgrade` into the new versioned data directory. No manual
-dump/restore needed.
-
-`pg_upgrade` must run as the old cluster's bootstrap superuser (its "install
-user"). The recipe uses `POSTGRES_USER`, which defaults to `postgres` — the right
-value for fresh installs and for clusters that came from the old bitnami recipe.
-If your cluster was bootstrapped with a different superuser (e.g. `discourse`),
-set `POSTGRES_USER` in the app `.env` before upgrading, otherwise `pg_upgrade`
-will refuse with an install-user mismatch.
+binaries and runs `pg_upgrade` into the new versioned data directory. The recipe
+adds a small entrypoint wrapper that injects the password secret and detects the
+old cluster's real install superuser (oid 10), so the upgrade works whether that
+user is `postgres` or `discourse`. No manual dump/restore needed.
 
 [`discourse/postgres`]: https://github.com/discourse/discourse-postgres
 

abra.sh

@@ -1,4 +1,4 @@
-export DB_ENTRYPOINT_VERSION=v7
+export DB_ENTRYPOINT_VERSION=v6
 export PG_BACKUP_VERSION=v3
 export APP_ENTRYPOINT_VERSION=v2
 export APP_INSTALL_SSL_VERSION=v1

cc-db-entrypoint.sh

@@ -1,20 +1,87 @@
 #!/bin/bash
 # Co-op Cloud entrypoint wrapper for the discourse/postgres image.
 #
-# The image (https://github.com/discourse/discourse-postgres) handles everything
-# itself, including the in-place major-version pg_upgrade on boot. The only thing
-# it can't do for us is read the docker secret: it expects the DB password in the
-# DB_PASSWORD / POSTGRES_PASSWORD env vars (no *_FILE support), so inject it from
-# /run/secrets here, then hand off to the image's real entrypoint.
+# discourse/postgres (https://github.com/discourse/discourse-postgres) is pgvector
+# plus a management layer that auto-upgrades an older cluster on boot. It does the
+# heavy lifting (apt-installs the old binaries, runs pg_upgrade, writes the new
+# cluster to the versioned PGDATA). This wrapper only fills the two gaps it leaves:
 #
-# The install user and data-checksum settings the upgrade needs are passed as
-# plain env vars in compose.yml (POSTGRES_USER, POSTGRES_INITDB_ARGS).
+#   1. Secrets: the image reads DB_PASSWORD / POSTGRES_PASSWORD from the process
+#      env (no *_FILE support), so inject them from the docker secret.
+#   2. Install user: the image runs `pg_upgrade --username="$POSTGRES_USER"` and
+#      initdb's the new cluster with $POSTGRES_USER, but never detects the OLD
+#      cluster's real bootstrap superuser (the install user, oid 10). pg_upgrade
+#      aborts unless the new cluster's install-user name matches the old one. Real
+#      deployments differ: some were bootstrapped with `postgres` as the install
+#      user (+ a separate `discourse` app role), others with `discourse` itself.
+#      So detect oid 10 from the old cluster and export POSTGRES_USER to match
+#      before handing off to the image's run-postgres.sh.
 set -e
 
+# --- 1. secret injection -----------------------------------------------------
 if [ -f /run/secrets/db_password ]; then
-  DB_PASSWORD="$(cat /run/secrets/db_password)"
-  export DB_PASSWORD
-  export POSTGRES_PASSWORD="$DB_PASSWORD"
+  pw="$(cat /run/secrets/db_password)"
+  export DB_PASSWORD="$pw"
+  export POSTGRES_PASSWORD="$pw"
+fi
+
+# --- 2. install-user detection (only matters on the upgrade path) ------------
+NEW_MAJOR="$(postgres --version | sed -rn 's/^[^0-9]*([0-9]+).*/\1/p')"
+
+# Newest existing cluster under the data mount (same search the image uses).
+PGVER_FILE="$(find /var/lib/postgresql -maxdepth 3 -type f -name PG_VERSION 2>/dev/null \
+  | xargs -I{} sh -c 'printf "%s " "{}"; cat "{}"' \
+  | sort -nk2,2 | tail -n1 | awk '{print $1}')"
+
+if [ -n "$PGVER_FILE" ]; then
+  OLD_DIR="$(dirname "$PGVER_FILE")"
+  OLD_MAJOR="$(cat "$PGVER_FILE")"
+  if [ "$OLD_MAJOR" != "$NEW_MAJOR" ]; then
+    echo "cc-db-entrypoint: existing pg${OLD_MAJOR} cluster at ${OLD_DIR}, image is pg${NEW_MAJOR} -> detecting install user"
+    OLD_BIN="/usr/lib/postgresql/${OLD_MAJOR}/bin"
+    if [ ! -x "$OLD_BIN/pg_ctl" ]; then
+      echo "cc-db-entrypoint: installing postgresql-${OLD_MAJOR} to read the old cluster"
+      apt-get update
+      apt-get install -y --no-install-recommends "postgresql-${OLD_MAJOR}" >/dev/null
+    fi
+    chown -R postgres "$OLD_DIR" 2>/dev/null || true
+
+    # Briefly start the old cluster on a local socket only, ask it for oid 10.
+    gosu postgres "$OLD_BIN/pg_ctl" -D "$OLD_DIR" -w \
+      -o "-c listen_addresses= -c unix_socket_directories=/tmp" start >/dev/null 2>&1 || true
+    detected=""
+    for login_role in discourse postgres; do
+      detected="$(gosu postgres psql -h /tmp -U "$login_role" -d postgres -tAc \
+        'select rolname from pg_roles where oid = 10' 2>/dev/null | tr -d '[:space:]')"
+      [ -n "$detected" ] && break
+    done
+    gosu postgres "$OLD_BIN/pg_ctl" -D "$OLD_DIR" -w stop >/dev/null 2>&1 || true
+
+    if [ -n "$detected" ]; then
+      echo "cc-db-entrypoint: old cluster install user is '$detected' -> POSTGRES_USER=$detected"
+      export POSTGRES_USER="$detected"
+    else
+      echo "cc-db-entrypoint: WARNING could not detect old install user; leaving POSTGRES_USER=${POSTGRES_USER:-<image default>}"
+    fi
+
+    # pg_upgrade refuses to run if the old and new clusters disagree on data
+    # checksums. PostgreSQL 18's initdb enables checksums by default, but the
+    # older clusters in this recipe's lineage (pg13-17) were created without
+    # them, so initdb the new cluster to match. Default to OFF (the lineage
+    # reality) and only enable when the old cluster positively reports them on.
+    csum=""
+    if [ -x "$OLD_BIN/pg_controldata" ]; then
+      csum="$("$OLD_BIN/pg_controldata" "$OLD_DIR" 2>/dev/null \
+        | awk -F: '/checksum version/{gsub(/[^0-9]/,"",$2); print $2}')"
+    fi
+    if [ "$csum" = "1" ]; then
+      echo "cc-db-entrypoint: old cluster data checksums ON -> initdb new cluster --data-checksums"
+      export POSTGRES_INITDB_ARGS="${POSTGRES_INITDB_ARGS:+$POSTGRES_INITDB_ARGS }--data-checksums"
+    else
+      echo "cc-db-entrypoint: old cluster data checksums OFF (version='${csum:-unknown}') -> initdb new cluster --no-data-checksums"
+      export POSTGRES_INITDB_ARGS="${POSTGRES_INITDB_ARGS:+$POSTGRES_INITDB_ARGS }--no-data-checksums"
+    fi
+  fi
 fi
 
 exec run-postgres.sh postgres

compose.yml

@@ -65,8 +65,8 @@ services:
   db:
     # discourse/postgres = pgvector + discourse's postgres management layer, which
     # auto-upgrades an older cluster in place on boot (pg_upgrade into the versioned
-    # PGDATA /var/lib/postgresql/${MAJOR}/docker). The cc-db-entrypoint wrapper only
-    # injects the password secret; everything else is driven by the env below.
+    # PGDATA /var/lib/postgresql/${MAJOR}/docker). The cc-db-entrypoint wrapper
+    # injects the password secret and detects the old cluster's install user.
     image: discourse/postgres:pg18
     networks:
       - internal
@@ -90,21 +90,13 @@ services:
       - POSTGRES_HOST_AUTH_METHOD=trust
       - POSTGRES_DB=discourse
       - DB_USER=discourse
-      # pg_upgrade runs as this role and initdb's the new cluster with it; it must
-      # match the OLD cluster's bootstrap superuser (oid 10). The image default
-      # `postgres` matches fresh installs and bitnami-origin clusters. Override in
-      # the app .env (POSTGRES_USER=...) only for a cluster bootstrapped differently.
-      - POSTGRES_USER=${POSTGRES_USER:-postgres}
-      # pg18's initdb enables data checksums by default, but pg13-17 clusters here
-      # have them off and pg_upgrade requires a match -> initialise without them.
-      - POSTGRES_INITDB_ARGS=--no-data-checksums
     healthcheck:
       test: "pg_isready -U discourse -d discourse"
       interval: 30s
       timeout: 10s
       retries: 5
-      # generous: on a major-version bump the image installs the old binaries and
-      # runs pg_upgrade on first boot before the server accepts connections —
+      # generous: a postgres major-version upgrade (apt install old binaries +
+      # pg_upgrade) runs in the entrypoint before the server accepts connections —
       # don't let the healthcheck kill an in-progress migration
       start_period: 15m
     deploy: