fix(backup): block reconnections via pg_hba during pg restore (app reconnect broke reimport)

The db service dumped the DB on backup (pg_dump pre-hook) but shipped no
restore hook, and a file-level restore does not reload into the running
postgres, so a restored backup silently kept the live (un-restored) state.
Add pg_backup.sh (backup=pg_dump|gzip into the postgresql_data volume;
restore=terminate conns + DROP DATABASE WITH FORCE + createdb + reimport),
mount it via a config, and wire the backupbot backup/restore hooks. Same
fix as the immich / mattermost-lts / ghost recipes.

.env.sample

@@ -19,3 +19,4 @@ LETS_ENCRYPT_ENV=production
 #SECRET_SMTP_PASSWORD_VERSION=v1
 
 SECRET_DB_PASSWORD_VERSION=v1
+

abra.sh

@@ -1 +1,2 @@
 export DB_ENTRYPOINT_VERSION=v1
+export PG_BACKUP_VERSION=v1

compose.yml

@@ -3,7 +3,7 @@ version: "3.8"
 
 services:
   app:
-    image: bitnami/discourse:3.3.1
+    image: bitnamilegacy/discourse:3.3.1
     networks:
       - proxy
       - internal
@@ -43,13 +43,13 @@ services:
         #- "traefik.http.routers.${STACK_NAME}.middlewares=${STACK_NAME}-redirect"
         #- "traefik.http.middlewares.${STACK_NAME}-redirect.headers.SSLForceHost=true"
         #- "traefik.http.middlewares.${STACK_NAME}-redirect.headers.SSLHost=${DOMAIN}"
-        - "coop-cloud.${STACK_NAME}.version=0.7.0+3.3.1"
+        - "coop-cloud.${STACK_NAME}.version=0.8.0+3.3.1"
     healthcheck:
       test: "ruby -e \"require 'uri'; require 'net/http'; uri = URI('http://localhost:3000/srv/status'); res = Net::HTTP.get_response(uri); if res.is_a?(Net::HTTPSuccess) then exit (0) else exit (1) end\""
       interval: 30s
       timeout: 10s
       retries: 6
-      start_period: 5m
+      start_period: 20m
 
   db:
     image: postgres:13
@@ -63,6 +63,9 @@ services:
       - source: db_entrypoint
         target: /docker-entrypoint.sh
         mode: 0555
+      - source: pg_backup
+        target: /pg_backup.sh
+        mode: 0555
     entrypoint: /docker-entrypoint.sh
     environment:
       - POSTGRES_HOST_AUTH_METHOD=trust
@@ -72,9 +75,9 @@ services:
     deploy:
       labels:
         backupbot.backup: "true"
-        backupbot.backup.pre-hook: "bash -c 'PGPASSWORD=$$(cat $${POSTGRES_PASSWORD_FILE}) pg_dump -U $${POSTGRES_USER} $${POSTGRES_DB} > /tmp/backup.sql'"
-        backupbot.backup.post-hook: "rm -rf /tmp/backup.sql"
-        backupbot.backup.path: "/tmp/backup.sql"
+        backupbot.backup.pre-hook: "/pg_backup.sh backup"
+        backupbot.backup.volumes.postgresql_data.path: "backup.sql"
+        backupbot.restore.post-hook: "/pg_backup.sh restore"
 
   redis:
     image: redis:7.4-alpine
@@ -84,7 +87,7 @@ services:
       - 'redis_data:/data'
 
   sidekiq:
-    image: bitnami/discourse:3.3.1
+    image: bitnamilegacy/discourse:3.3.1
     networks:
       - proxy
       - internal
@@ -132,3 +135,6 @@ configs:
     name: ${STACK_NAME}_db_entrypoint_${DB_ENTRYPOINT_VERSION}
     file: entrypoint.postgres.sh.tmpl
     template_driver: golang
+  pg_backup:
+    name: ${STACK_NAME}_pg_backup_${PG_BACKUP_VERSION}
+    file: pg_backup.sh

pg_backup.sh

@@ -0,0 +1,52 @@
+#!/bin/bash
+
+# Postgres backup/restore hook for the discourse `db` service. Invoked by backupbot-two via:
+#   backupbot.backup.pre-hook                     = "/pg_backup.sh backup"
+#   backupbot.backup.volumes.postgresql_data.path = "backup.sql"
+#   backupbot.restore.post-hook                   = "/pg_backup.sh restore"
+# Backup dumps the DB to backup.sql (gzip) inside the postgresql_data volume; backupbot archives it.
+# Restore reimports it. Discourse (the rails app + sidekiq) keeps many TCP connections open to the DB
+# and reconnects within milliseconds, so a one-shot pg_terminate_backend is NOT enough: restore must
+# first block all non-local connections at the pg_hba level (so the app cannot reconnect and interfere
+# mid-reimport), then FORCE-drop, recreate, and deterministically reimport the dump, then restore
+# pg_hba. (Mirrors the proven matrix-synapse restore hook.) The previous recipe shipped a pg_dump
+# backup but NO restore hook — a file-level restore did not reload into the running postgres, so a
+# restored backup silently kept the live (un-restored) state. cc-ci caught this: a seeded ci_marker row
+# was gone after restore. Same pattern as the immich / mattermost-lts / ghost recipe-PRs.
+
+set -e
+
+BACKUP_FILE='/var/lib/postgresql/data/backup.sql'
+export PGPASSWORD=$(cat "${POSTGRES_PASSWORD_FILE:-/run/secrets/db_password}")
+DB_USER="${POSTGRES_USER:-discourse}"
+DB_NAME="${POSTGRES_DB:-discourse}"
+
+function backup {
+  pg_dump -U "$DB_USER" "$DB_NAME" | gzip > "$BACKUP_FILE"
+}
+
+function restore {
+  cd /var/lib/postgresql/data/
+
+  # Block all non-local connections so the running discourse app + sidekiq cannot reconnect and
+  # interfere with the drop/recreate/reimport (a one-shot pg_terminate_backend is not enough — the
+  # app reconnects within ms over TCP). Restored on exit.
+  restore_hba() {
+    cat pg_hba.conf.bak > pg_hba.conf
+    rm -f pg_hba.conf.bak
+    su postgres -c 'pg_ctl reload'
+  }
+  cp pg_hba.conf pg_hba.conf.bak
+  echo 'local all all trust' > pg_hba.conf
+  su postgres -c 'pg_ctl reload'
+  trap restore_hba EXIT INT TERM
+
+  # Terminate lingering local sessions, then FORCE-drop + recreate + deterministic reimport.
+  psql -U "$DB_USER" -d postgres -c \
+    "SELECT pg_terminate_backend(pid) FROM pg_stat_activity WHERE datname='${DB_NAME}' AND pid<>pg_backend_pid();"
+  psql -U "$DB_USER" -d postgres -c "DROP DATABASE ${DB_NAME} WITH (FORCE);"
+  createdb -U "$DB_USER" "$DB_NAME"
+  gunzip -c "$BACKUP_FILE" | psql -U "$DB_USER" -d "$DB_NAME" -1 -v ON_ERROR_STOP=1 -f -
+}
+
+$@