fix(backup): add pg_backup.sh + proper backup/restore hooks, 20m start_period

This commit is contained in:
notplants
2026-06-02 07:50:05 +00:00
committed by notplants
parent 0f873433ba
commit ec7bbdf786
3 changed files with 63 additions and 4 deletions

View File

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

View File

@ -49,7 +49,7 @@ services:
interval: 30s
timeout: 10s
retries: 6
start_period: 5m
start_period: 20m
db:
image: postgres:16
@ -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
@ -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

52
pg_backup.sh Executable file
View File

@ -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 -
}
$@