gendesign/ops/backup.sh
bot-backend 1cb14c6a35
All checks were successful
CI Trade-In / changes (pull_request) Successful in 7s
CI Trade-In / backend-tests (pull_request) Has been skipped
CI Trade-In / browser-tests (pull_request) Has been skipped
CI Trade-In / frontend-checks (pull_request) Has been skipped
CI / changes (pull_request) Successful in 8s
CI / frontend-tests (pull_request) Has been skipped
CI / backend-tests (pull_request) Has been skipped
CI / openapi-codegen-check (pull_request) Has been skipped
fix(ops): бэкап без выгрузки в S3 падает, а не рапортует успех
Проверка хостов за шесть дней до переезда показала, что на Poincare пустой
crontab: ни бэкапов, ни сторожей, ни конфигов. Сегодня это неважно — базы ещё
на Beget, где всё работает. Тридцатого августа они переезжают, и бэкапы там
просто не начнутся.

Хуже, что отказ был бы тихим. Выгрузка в S3 была опциональной: без четырёх
переменных скрипт писал "backup stays local only" и выходил с кодом ноль.
Не ошибка, не алерт — успешный прогон. Сторож свежести на новом хосте тоже
не установлен, так что промолчали бы оба.

Теперь отсутствие любой из S3_ENDPOINT / S3_BUCKET / S3_ACCESS_KEY /
S3_SECRET_KEY роняет прогон с ненулевым кодом — до снятия дампа, потому что
незачем тратить время и место на дамп, которому некуда уехать. Образец
поведения взят из backup-forgejo.sh, где так было с самого начала.

BACKUP_ALLOW_LOCAL_ONLY=1 оставляет прежний путь для машины без ключей.

Ревью показало, что первая версия этой форточки возвращала ровно ту дыру,
которую чинит PR: локальный прогон всё равно писал сентинел и выходил нулём,
поэтому сторож свежести оставался зелёным, а единственным следом был WARNING
в логе, который никто не читает. Достаточно было раскомментировать одну
строку в /etc/default — и «настроено» выглядело бы неотличимо от рабочего.
Теперь такой прогон сентинел НЕ пишет и сразу зовёт notify().

Попутно найден живой баг: в lib-backup.sh внутри блока после ||
литеральное `\n` разбиралось шеллом как команда `n`, поэтому запасной канал
оповещения по почте не срабатывал никогда. То есть при недоступности Telegram
мы теряли уведомления молча — тот же класс тишины.

restore-drill.sh поставлен в расписание Poincare: он существовал, но не
запускался нигде, и восстановимость дампов автоматически не проверялась ни
разу. gzip -t и трейлер pg_dump говорят «файл не битый», а это не то же
самое, что «база поднимается». У самой дрели своего сторожа нет и MAILTO в
cron не задан, поэтому её падение теперь тоже уходит в notify(). Заодно
docker rm -f получил -v: регулярный запуск иначе оставлял бы dangling-том
размером с базу до воскресной чистки.

Имя бакета в примере исправлено на gendsgn-backups — на хостах настроен
именно он, forgejo пишет туда же под префиксом forgejo/.

Две находки ревью намеренно не закрыты. Фолбэк на общий конфиг в
tradein-скрипте остался elif: если появится /etc/default/tradein-backup без
S3-переменных, прогон упадёт вместо чтения общего файла — под новой политикой
это и есть желаемое поведение. Дрель проверяет только серию gendesign:
restore-drill.sh не умеет батч, а расширять его в этом PR значит смешивать
задачи.

Хостовая часть — установка crontab и конфига на Poincare — в PR не входит.
Конфиг уже положен вручную и выгрузка с Poincare проверена живым объектом;
crontab ставится в окно переезда, потому что продуктового postgres там пока
нет и ночные прогоны падали бы впустую.

Refs #3085, #3008
2026-08-24 20:48:45 +03:00

252 lines
13 KiB
Bash
Executable file

#!/usr/bin/env bash
# Daily Postgres backup for the MAIN gendesign DB. Runs from cron on the prod VM.
#
# Modeled on the proven ops/.../backup-tradein-db.sh (#397). Hardened for #71
# after the main-DB backup silently broke (last good dump 2026-05-27): a raw
# deploy-time hard reset of the repo checkout onto origin/main reset this
# file's mode to 644, so a cron entry that invoked the raw path got
# "Permission denied" every run.
#
# Robustness measures here (extended for #2203):
# - cron should invoke via `bash <path>` so a missing +x bit can't break it
# (this file is ALSO committed 100755, and deploy.yml re-chmods ops/*.sh);
# - sanity-check: a suspiciously small dump (< MIN_DUMP_BYTES) is treated as a
# failed dump — it's deleted and the script exits non-zero, so a good prior
# dump is never pruned in favour of a truncated one;
# - integrity-check: gzip -t + a trailer sentinel catch a truncated dump that
# is neither empty nor undersized (see verify_dump_integrity below);
# - globals (roles/GRANTs) are dumped separately via `pg_dumpall --globals-only`
# — `pg_dump` never includes these, so without this a restore has tables
# but no owning roles/privileges;
# - retention: keep the KEEP most-recent local dumps, delete the rest — main
# dumps and globals dumps are tracked as separate series.
#
# Usage (cron — note `bash`, not a bare path, so +x is irrelevant):
# 30 3 * * * bash /opt/gendesign/ops/backup.sh >> /var/log/gendesign-backup.log 2>&1
#
# S3 upload — REQUIRED by default (#3085). Set these in /etc/default/gendesign-backup
# (root-owned, chmod 600, NOT in git). Without all four, the script fails
# LOUDLY before taking a dump — set BACKUP_ALLOW_LOCAL_ONLY=1 to explicitly
# accept a local-only backup instead (e.g. local testing):
# S3_ENDPOINT=https://s3.ru-1.storage.selcloud.ru # Selectel S3
# S3_BUCKET=gendsgn-backups
# S3_ACCESS_KEY=...
# S3_SECRET_KEY=...
# A redacted template lives at ops/gendesign-backup.default.example.
#
# Restore: see ops/restore.sh (destructive, INTO the live DB).
# Restore drill (safe, throwaway container): see ops/restore-drill.sh.
#
# Missed-run detection (#2203): on a verified-good run this writes a sentinel
# file (SENTINEL_FILE) with the UTC timestamp. A SEPARATE cron entry running
# ops/check-backup-staleness.sh against that file is what actually alerts if
# cron stops firing — see that script's header for the alert channel and
# cron line. This script itself does NOT alert; it only ever marks success.
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# shellcheck source=./lib-backup.sh
source "$SCRIPT_DIR/lib-backup.sh"
# --- config (env-overridable) ---
COMPOSE_DIR="${COMPOSE_DIR:-/opt/gendesign}"
COMPOSE_FILE="${COMPOSE_FILE:-docker-compose.prod.yml}"
COMPOSE_PROJECT="${COMPOSE_PROJECT:-gendesign}"
LOCAL_BACKUP_DIR="${LOCAL_BACKUP_DIR:-/opt/gendesign/backups}"
KEEP="${KEEP:-7}" # how many recent local dumps to keep
MIN_DUMP_BYTES="${MIN_DUMP_BYTES:-51200}" # 50 KiB floor; gzip'd schema-only dump
# is already > this, so a healthy dump
# never trips it. Real DB is far larger.
# (No equivalent floor for the globals
# dump — a handful of roles legitimately
# gzips to well under this.)
SENTINEL_FILE="${SENTINEL_FILE:-${LOCAL_BACKUP_DIR}/.last_success}"
# Optional S3 env. Loaded from /etc/default/gendesign-backup if present.
[[ -f /etc/default/gendesign-backup ]] && source /etc/default/gendesign-backup
# --- guard: S3 upload is REQUIRED by default (#3085). A backup that only
# lives on the same disk as the DB it protects is not a backup once
# Poincare has no configured fallback host to copy from — before this
# guard, a missing/incomplete /etc/default/gendesign-backup made the
# script log "backup stays local only" and exit 0, indistinguishable
# from a real off-box backup in the cron log. Fail LOUD, before the dump
# is even taken (no point spending time/disk on a dump that can't leave
# the box). BACKUP_ALLOW_LOCAL_ONLY=1 is the explicit escape hatch for
# running on a box without S3 creds on purpose (e.g. local testing) —
# it still warns loudly so it can't be forgotten silently. ---
if [[ -z "${S3_ENDPOINT:-}" || -z "${S3_BUCKET:-}" || -z "${S3_ACCESS_KEY:-}" || -z "${S3_SECRET_KEY:-}" ]]; then
if [[ "${BACKUP_ALLOW_LOCAL_ONLY:-0}" == "1" ]]; then
log "WARNING: S3 vars not fully set — BACKUP_ALLOW_LOCAL_ONLY=1, proceeding with a LOCAL-ONLY backup. This copy will NOT survive loss of this VM." >&2
BACKUP_LOCAL_ONLY_USED=1
else
log "ERROR: S3 upload is not configured — missing one or more of S3_ENDPOINT / S3_BUCKET / S3_ACCESS_KEY / S3_SECRET_KEY." >&2
log "ERROR: fill them into /etc/default/gendesign-backup (see ops/gendesign-backup.default.example)," >&2
log "ERROR: or set BACKUP_ALLOW_LOCAL_ONLY=1 to explicitly accept a local-only backup." >&2
exit 1
fi
fi
# Integrity check beyond "non-empty"/"big enough": a dump truncated mid-write
# (disk full, OOM-kill, docker exec dropped) can still gzip into a structurally
# valid, non-tiny .gz — neither `-s` nor MIN_DUMP_BYTES catch that reliably.
# Two checks, cheapest first:
# 1. gzip -t — catches a corrupted/truncated gzip stream itself.
# 2. trailer — pg_dump/pg_dumpall always write a fixed "I finished writing"
# comment as literally the last line of the stream; a dump cut off
# mid-write is missing it even when the gzip framing looks fine.
# Checked via `gunzip -c | tail -N` so we never materialize the full
# decompressed dump on disk just to look at its last few lines.
verify_dump_integrity() {
local file="$1" trailer="$2" label="$3"
if ! gzip -t "$file" 2>/dev/null; then
log "ERROR: ${label} failed gzip integrity check -> $file" >&2
return 1
fi
if ! gunzip -c "$file" 2>/dev/null | tail -5 | grep -qF -- "$trailer"; then
log "ERROR: ${label} missing trailer '${trailer}' -> $file (truncated dump?)" >&2
return 1
fi
return 0
}
# --- run ---
mkdir -p "$LOCAL_BACKUP_DIR"
# Underscore naming kept (gendesign_YYYYMMDD_HHMMSS.sql.gz) so this matches
# pre-existing prod dumps and restore.sh's example path. UTC for stable ordering.
ts=$(date -u +'%Y%m%d_%H%M%S')
out="${LOCAL_BACKUP_DIR}/gendesign_${ts}.sql.gz"
globals_out="${LOCAL_BACKUP_DIR}/gendesign_globals_${ts}.sql.gz"
cd "$COMPOSE_DIR"
compose() { docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" "$@"; }
# Read DB user/name from the running compose env; fall back to "gendesign".
DB_USER=$(compose exec -T postgres printenv POSTGRES_USER 2>/dev/null | tr -d '\r' || true)
DB_NAME=$(compose exec -T postgres printenv POSTGRES_DB 2>/dev/null | tr -d '\r' || true)
DB_USER=${DB_USER:-gendesign}
DB_NAME=${DB_NAME:-gendesign}
log "Dumping ${DB_NAME} as ${DB_USER} -> ${out}"
# --clean --if-exists -> dump is self-sufficient for restore from scratch.
# --no-owner -> restore does not require identical roles.
# `set -o pipefail` (from set -euo pipefail) makes a pg_dump failure fail the
# whole pipe, so a Postgres error can't yield a "successful" tiny gzip. Stderr
# is NOT redirected to /dev/null (was until #2203) — a pg_dump error needs to
# land in the cron log, not get silently swallowed.
compose exec -T postgres \
pg_dump -U "$DB_USER" -d "$DB_NAME" --no-owner --clean --if-exists \
| gzip -9 > "$out"
# --- sanity-check: refuse to keep (and thus never prune good dumps for) a
# truncated/empty dump. ---
if [[ ! -s "$out" ]]; then
log "ERROR: dump is empty -> $out — removing, keeping previous good dumps." >&2
rm -f "$out"
exit 1
fi
dump_bytes=$(wc -c < "$out" | tr -d ' ')
if (( dump_bytes < MIN_DUMP_BYTES )); then
log "ERROR: dump only ${dump_bytes} bytes (< ${MIN_DUMP_BYTES} floor) — likely a failed dump." >&2
log "Removing suspect dump, NOT pruning older good dumps." >&2
rm -f "$out"
exit 1
fi
if ! verify_dump_integrity "$out" "-- PostgreSQL database dump complete" "main dump"; then
log "Removing suspect dump, NOT pruning older good dumps." >&2
rm -f "$out"
exit 1
fi
log "Dump OK: ${out} ($(du -h "$out" | cut -f1), ${dump_bytes} bytes)"
# --- globals (roles/GRANTs) — pg_dump never includes these; without them a
# restored dump has tables and data but no owning roles/privileges. ---
log "Dumping globals (roles/GRANTs) -> ${globals_out}"
compose exec -T postgres \
pg_dumpall -U "$DB_USER" --globals-only \
| gzip -9 > "$globals_out"
if [[ ! -s "$globals_out" ]]; then
log "ERROR: globals dump is empty -> $globals_out — removing." >&2
rm -f "$globals_out"
exit 1
fi
if ! verify_dump_integrity "$globals_out" "-- PostgreSQL database cluster dump complete" "globals dump"; then
rm -f "$globals_out"
exit 1
fi
log "Globals dump OK: ${globals_out} ($(du -h "$globals_out" | cut -f1))"
# --- S3 upload (only if all four vars present — guaranteed unless the
# BACKUP_ALLOW_LOCAL_ONLY escape hatch above was used) — both the main dump
# and the globals dump go up, same bucket, same layout. ---
if [[ -n "${S3_ENDPOINT:-}" && -n "${S3_BUCKET:-}" && -n "${S3_ACCESS_KEY:-}" && -n "${S3_SECRET_KEY:-}" ]]; then
for f in "$out" "$globals_out"; do
log "Uploading to s3://${S3_BUCKET}/$(basename "$f")"
# aws-cli v2 ships its own CA bundle (baked into botocore) instead of
# trusting the system store, and it's missing the root that Selectel's
# cert chains up to — so uploads fail with CERTIFICATE_VERIFY_FAILED
# unless we point it at the container's system store, which has it.
docker run --rm \
-e AWS_ACCESS_KEY_ID="$S3_ACCESS_KEY" \
-e AWS_SECRET_ACCESS_KEY="$S3_SECRET_KEY" \
-e AWS_CA_BUNDLE=/etc/ssl/certs/ca-certificates.crt \
-v "$LOCAL_BACKUP_DIR":/backup:ro \
amazon/aws-cli:latest \
--endpoint-url "$S3_ENDPOINT" \
s3 cp --no-progress "/backup/$(basename "$f")" "s3://${S3_BUCKET}/"
done
log "S3 upload OK"
else
log "S3 vars not set — backup stays local only"
fi
# --- retention: keep the KEEP most-recent local dumps, delete the rest.
# Runs only AFTER a verified-good dump above, so a failed run (which exits
# early) can never delete older good dumps. Main dumps and globals dumps
# are separate series (the `[0-9]` after the shared `gendesign_` prefix
# keeps the globals_* files, which share that prefix, out of this glob). ---
# SC2012: ls is fine here — filenames are fully controlled (gendesign_<ts>.sql.gz,
# no spaces/newlines) and we need ls's -t mtime sort for "keep newest N" (same
# pattern as the proven backup-tradein-db.sh).
# shellcheck disable=SC2012
ls -1t "$LOCAL_BACKUP_DIR"/gendesign_[0-9]*.sql.gz 2>/dev/null \
| tail -n +"$((KEEP + 1))" \
| xargs -r rm -f
# shellcheck disable=SC2012
ls -1t "$LOCAL_BACKUP_DIR"/gendesign_globals_*.sql.gz 2>/dev/null \
| tail -n +"$((KEEP + 1))" \
| xargs -r rm -f
# Count via glob arrays (no ls parsing). nullglob -> empty array if no match.
shopt -s nullglob
remaining=( "$LOCAL_BACKUP_DIR"/gendesign_[0-9]*.sql.gz )
remaining_globals=( "$LOCAL_BACKUP_DIR"/gendesign_globals_*.sql.gz )
shopt -u nullglob
log "Backup done. Local dumps retained: ${#remaining[@]} data + ${#remaining_globals[@]} globals (KEEP=${KEEP})."
# --- missed-run detection (#2203): mark success ONLY here, after every
# guard above (dump/globals integrity, retention, and — since this whole
# script runs under `set -euo pipefail` — any S3 upload) has passed. See
# ops/check-backup-staleness.sh for the piece that actually alerts if
# this stops being refreshed. A LOCAL-ONLY run (BACKUP_ALLOW_LOCAL_ONLY=1
# escape hatch above) does NOT count as a verified backup for this
# purpose — writing the sentinel here would leave the staleness watchdog
# permanently green while the box has no off-box copy at all (#3085
# reopened via the escape hatch). Instead: skip the sentinel (watchdog
# alerts within max_age_hours) and notify() immediately so it isn't only
# a WARNING buried in a log nobody tails. ---
if [[ "${BACKUP_LOCAL_ONLY_USED:-0}" == "1" ]]; then
log "Skipping sentinel write — this was a LOCAL-ONLY backup (BACKUP_ALLOW_LOCAL_ONLY=1), not a verified off-box backup." >&2
notify "⚠️ gendesign main backup: LOCAL-ONLY run (S3 not configured, BACKUP_ALLOW_LOCAL_ONLY=1). Dump ${out} is NOT off-box. Sentinel NOT updated — staleness watchdog will alert if this isn't fixed."
else
write_sentinel "$SENTINEL_FILE"
fi