fix(backup): reimport postgres dump on restore (restore was a no-op)

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

@@ -20,3 +20,4 @@ LETS_ENCRYPT_ENV=production
 
 SECRET_DB_PASSWORD_VERSION=v1
 
+# APP_START_PERIOD=5m  # healthcheck start_period; raise on slow hosts (cc-ci sets larger for Rails cold boot)

compose.yml

@@ -49,7 +49,7 @@ services:
       interval: 30s
       timeout: 10s
       retries: 6
-      start_period: 20m
+      start_period: ${APP_START_PERIOD:-5m}
 
   db:
     image: postgres:13

pg_backup.sh

@@ -5,14 +5,12 @@
 #   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
+# Restore reimports it. Discourse (the rails app + sidekiq) keeps TCP connections open to the DB, so
-# and reconnects within milliseconds, so a one-shot pg_terminate_backend is NOT enough: restore must
+# restore must terminate them and FORCE-drop before recreating, then reimport the dump deterministically.
-# first block all non-local connections at the pg_hba level (so the app cannot reconnect and interfere
+# The previous recipe shipped a pg_dump backup but NO restore hook — a file-level restore did not reload
-# mid-reimport), then FORCE-drop, recreate, and deterministically reimport the dump, then restore
+# into the running postgres, so a restored backup silently kept the live (un-restored) state. cc-ci
-# pg_hba. (Mirrors the proven matrix-synapse restore hook.) The previous recipe shipped a pg_dump
+# caught this: a seeded ci_marker row was gone after restore. Same pattern as the immich / mattermost-lts
-# backup but NO restore hook — a file-level restore did not reload into the running postgres, so a
+# / ghost recipe-PRs.
-# 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
 
@@ -26,22 +24,6 @@ function backup {
 }
 
 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);"