lyraflow/backup-lib.sh at main · lyraflow/lyraflow

History

702 lines (666 loc) · 32.9 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

490

491

492

493

494

495

496

497

498

499

500

501

502

503

504

505

506

507

508

509

510

511

512

513

514

515

516

517

518

519

520

521

522

523

524

525

526

527

528

529

530

531

532

533

534

535

536

537

538

539

540

541

542

543

544

545

546

547

548

549

550

551

552

553

554

555

556

557

558

559

560

561

562

563

564

565

566

567

568

569

570

571

572

573

574

575

576

577

578

579

580

581

582

583

584

585

586

587

588

589

590

591

592

593

594

595

596

597

598

599

600

601

602

603

604

605

606

607

608

609

610

611

612

613

614

615

616

617

618

619

620

621

622

623

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

# shellcheck shell=bash

# Shared helpers for backup.sh (and restore.sh, which sources this same file).

# Sourced, never executed, so it deliberately has no shebang and is not

# executable — the `shell=bash` directive above is what tells shellcheck which

# dialect to check it as.

# Targets bash 3.2, the version macOS still ships: no associative arrays, no

# `mapfile`, no `${var^^}`. An operator restoring at 2am is not going to

# upgrade their shell first.

# Every `docker` invocation here is plain `docker compose` with no `-f`, so

# COMPOSE_FILE and COMPOSE_PROJECT_NAME work the way an operator (and the

# round-trip tests) expect.

# Service names as they appear in docker-compose.yml, named once so the two

# scripts cannot drift apart.

APP_SERVICE=lyraflow

CH_SERVICE=clickhouse

PG_SERVICE=postgres

# The application database and role in each store.

CH_DATABASE=lyraflow

PG_DATABASE=lyraflow

PG_USER=lyraflow

# Where the ClickHouse `backups` disk is mounted inside its container. Declared

# by docker/clickhouse/backup-disk.xml; see that file for why the disk lives

# inside the existing data volume and what that obliges this script to clean up.

CH_BACKUP_DIR=/var/lib/clickhouse/backups

# Set to 1 immediately BEFORE `docker compose stop`, not after it. Read by

# start_app_if_stopped, which both the happy path and the EXIT trap call.

# Before, because `docker compose stop` can stop the container and still exit

# non-zero — Ctrl-C during the 30-second drain exits 130 with the container

# already down. Setting the flag afterwards means that path never sets it, the

# trap sees 0, and the app is left stopped with nothing announcing it.

# Measured, with a shim that really stopped the container and then reported

# failure: `exited/unhealthy` immediately after the script and still

# `exited/unhealthy` 50 seconds later. Setting it early is free — M2 in the

# mutation set established that `docker compose start` against a running

# container is a no-op that does not even move `StartedAt`.

APP_STOPPED=0

# Set to 1 only once `BACKUP` has reported success, so cleanup removes an

# archive this run created and never one belonging to a concurrent run. See

# remove_in_container_artefact.

CH_ARTEFACT_CREATED=0

# ---------------------------------------------------------------------------

# Output

# ---------------------------------------------------------------------------

# say <text…> progress, on stdout

# note <text…> diagnostics, on stderr

# NEITHER CAN FAIL, and that is the whole point of them existing.

# `./backup.sh DEST | head -1` closes this script's stdout as soon as the

# reader has what it wants, and a plain `echo` then dies of SIGPIPE. Measured

# before the fix:

# PIPESTATUS=141 head=0 <- 141 = 128 + SIGPIPE

# lyraflow-ci-lyraflow-1 exited Exited (0)

# --- full stderr of that run --- (nothing but docker's own two lines)

# The app was stopped there because start_app_if_stopped's FIRST STATEMENT was

# a write, so the signal landed before `docker compose start`. It was not, as

# an earlier version of this comment claimed, because the EXIT trap failed to

# run -- bash runs the EXIT trap on SIGPIPE death, measured on 5.2.21 and

# 5.3.9. Ordering is what keeps the app up.

# What these two buy is that the run continues rather than ending at its first

# message: `trap '' PIPE` in backup.sh turns the signal into a write error, and

# the `|| true` here turns the write error into nothing. Both are needed -- the

# trap alone leaves `echo` returning non-zero, which `set -e` turns into an

# aborted run by a different route.

# Reachable by `| less` and pressing q, by `| head -n N` in a wrapper capping a

# log, by `| grep -m1`, and by `| mail` or `| logger` where the reader dies.

# Note a closed stdout does NOT abandon the backup: a wrapper that caps its log

# asked for a shorter log, not for an unfinished backup.

# The `2>/dev/null` suppresses bash's own "echo: write error: Broken pipe"

# diagnostic, which otherwise replaces the message the reader did not want with

# a noisier one it also did not want. Redirections apply left to right, so

# note()'s fd1 is the real stderr before fd2 is discarded.

say() { echo "$@" 2>/dev/null || true; }

note() { echo "$@" >&2 2>/dev/null || true; }

# ---------------------------------------------------------------------------

# Failure reporting

# ---------------------------------------------------------------------------

# fail <step> <detail…>

# This script runs unattended at 4am and the log line is the only artefact of

# the failure, so it has to answer the two questions an operator has

# immediately: which step died, and whether their data is intact. A stack trace

# or a bare `Code: 598` answers neither.

# The "no data was changed" claim is true on every path because neither script

# path writes to either store: BACKUP, pg_dump and every count query are reads.

# The only thing written inside a container is the scratch archive on the

# ClickHouse backups disk, which is not data and which cleanup removes.

# Exits 1, which fires the EXIT trap — so the app is restarted after this

# prints, not before.

fail() {

local step="$1"

shift

note ""

note "ERROR: the backup failed during the ${step} step."

local line

for line in "$@"; do

note " ${line}"

done

note ""

note "No data was changed: this script only reads from ClickHouse and Postgres,"

note "so both stores still hold exactly what they held before it started."

exit 1

}

# ---------------------------------------------------------------------------

# Store access

# ---------------------------------------------------------------------------

# ch_query <sql> [database]

# The ClickHouse password is already in the container's own environment

# (CLICKHOUSE_PASSWORD, which compose sets from .env). It is referenced INSIDE

# the single-quoted `sh -c` program, so the host shell never expands it and the

# secret never reaches the host's argv — `ps auxww` during a multi-second

# BACKUP shows the literal string "$CLICKHOUSE_PASSWORD" and nothing else.

# Do NOT interpolate "$CLICKHOUSE_PASSWORD" from the host here.

# The database argument exists because the normal binding to `lyraflow` is

# unusable at exactly the moments restore.sh needs a client: the HTTP/native

# session sets its current database on connect, and naming one that does not

# exist is itself an error (Code: 81, UNKNOWN_DATABASE) — which is the state

# the database is in between a DROP and its RESTORE. Pass `default` there.

# `< /dev/null` is load-bearing, not decoration: `docker compose exec -T`

# forwards the caller's stdin into the container, so a ch_query inside a

# `while read` loop eats the rest of the loop's input. Observed exactly once,

# with ch_row_counts reporting the first table and silently dropping the other

# four — a manifest that looks well-formed and describes a quarter of the

# database. Every exec below closes stdin for the same reason.

ch_query() {

docker compose exec -T "$CH_SERVICE" sh -c \

'clickhouse-client --user "$CLICKHOUSE_USER" --password "$CLICKHOUSE_PASSWORD" \

--database "$1" --query "$0"' \

"$1" "${2:-$CH_DATABASE}" < /dev/null

}

# pg_query <sql>

# Same argument as ch_query for the password: PGPASSWORD is assigned inside the

# single-quoted program from the container's own POSTGRES_PASSWORD, so it never

# appears in the host's process list. The official image's pg_hba grants `local

# all all trust`, so this would work without a password today — it is written

# this way so it keeps working against a hardened pg_hba rather than failing at

# 4am the day someone tightens it.

pg_query() {

docker compose exec -T "$PG_SERVICE" sh -c \

'PGPASSWORD="$POSTGRES_PASSWORD" psql -U "$1" -d "$2" -At -v ON_ERROR_STOP=1 -c "$0"' \

"$1" "$PG_USER" "$PG_DATABASE" < /dev/null

}

# artefact_write <path> — THE ONLY PLACE IN EITHER SCRIPT THAT CREATES A FILE

# BY WRITING TO IT.

# The qualifier is not pedantry. `mv "$tmp" "$out/MANIFEST"` also brings a path

# into existence; it is safe because it renames a file this function already

# created, carrying that mode with it, and it is deliberately the only other

# way any path here appears. An absolute claim in this header would be false,

# and E3 below is what a false absolute costs.

# Reads stdin, writes <path>. Every artefact goes through here, which is what

# makes "created at 0600 by the process umask, never widened, never staged

# through anything looser" a property of one auditable line instead of a claim

# about the whole codebase.

# KNOWN LIMITATION (E3). The chokepoint is enforced by two audits — one static

# rule that this is the only data redirect, one runtime allow-list of the

# commands that actually run — and a *simplification* slips between them:

# replacing `| artefact_write "$tmp"` with `sort -o "$tmp"` creates the file

# with no redirect at all, using an already-allow-listed command, on a covered

# path. The mode stays 0600 because the umask still applies, so there is no

# security regression today; what is lost is the invariant, silently. Closing

# it needs an assertion on how many times artefact_write is *invoked* per run,

# which is left undone deliberately rather than unnoticed.

# This shape exists because two rounds of banning spellings lost. First `chmod`

# was banned, so the next attempt used `install -m 600`; that was banned too, so

# the next used `docker compose cp` into a scratch file and then `cat` that

# scratch through the umask into the real destination -- no chmod, no install,

# one correctly placed `umask 077`, both permission tests green, and a scratch

# file holding the plaintext Postgres password sitting at 0640 for the length of

# the copy. You cannot win a spelling war against file creation; you can only

# reduce the number of places it happens to one and audit that.

# The test enforces the chokepoint rather than the vocabulary: exactly one data

# redirect in either script and it is the one below, plus an allow-list of the

# external commands these scripts may invoke at all, so a new way to make a file

# cannot be introduced without failing a test whatever it is spelled.

# `cat` and not `cp`/`install`/`dd`: the caller supplies bytes on a pipe, so

# there is no source file whose mode could be copied along with them.

artefact_write() {

cat > "$1"

}

pg_dump_to() {

# A pipe, not a redirect. `set -o pipefail` is what makes the exit status

# still pg_dump's -- verified by the tests that inject a pg_dump failure and

# require the run to fail and leave nothing behind. restore.sh must set

# pipefail too if it uses this.

docker compose exec -T "$PG_SERVICE" sh -c \

'PGPASSWORD="$POSTGRES_PASSWORD" pg_dump -U "$0" -d "$1" -Fc' \

"$PG_USER" "$PG_DATABASE" < /dev/null | artefact_write "$1"

}

# copy_ch_artefact_to <file>

# `docker compose exec … cat > file`, deliberately, and NOT `docker compose cp`.

# `cp` reproduces the mode the file has inside the container and ignores the

# caller's umask: measured on this branch, the same archive arrived as 0640 via

# `cp` and 0600 via this redirect, byte-for-byte identical otherwise. 0640 is

# whatever ClickHouse happened to choose, not what this script chose, and the

# archive contains the identity dictionaries' Postgres password — see

# backup.sh's umask comment. A redirect is the only form the umask applies to.

# A pipe into the chokepoint, as pg_dump_to is: `set -o pipefail` is what keeps

# the exit status the copy's own, and restore.sh must set it too.

copy_ch_artefact_to() {

docker compose exec -T "$CH_SERVICE" cat "$CH_BACKUP_DIR/$CH_FILE" < /dev/null |

artefact_write "$1"

}

# ---------------------------------------------------------------------------

# Compose helpers

# ---------------------------------------------------------------------------

service_running() {

docker compose ps --status running --services 2>/dev/null | grep -qx "$1"

}

# service_image <service> — the image reference the container was created from.

# Recorded in the manifest because "which images produced this backup" is the

# first thing a restore needs and the last thing anyone writes down. Read from

# the container rather than from `docker compose config --images`, which does

# not say which image belongs to which service, and works while the container

# is stopped, which is when the manifest is written.

service_image() {

local cid

cid="$(docker compose ps -aq "$1" 2>/dev/null | head -n 1)"

if [ -z "$cid" ]; then

echo "unknown"

return 0

docker inspect --format '{{.Config.Image}}' "$cid" 2>/dev/null || echo "unknown"

}

# wait_until_stopped <service> <seconds>

# `docker compose stop` returning is not the same as the grace period having

# elapsed. The drain is the entire reason for quiescing, so poll rather than

# assume.

wait_until_stopped() {

local i=0

while [ "$i" -lt "$2" ]; do

service_running "$1" || return 0

i=$((i + 1))

sleep 1

done

return 1

}

# wait_until_healthy <service> <seconds>

wait_until_healthy() {

local cid i status

cid="$(docker compose ps -aq "$1" 2>/dev/null | head -n 1)"

[ -n "$cid" ] || return 1

i=0

while [ "$i" -lt "$2" ]; do

status="$(docker inspect \

--format '{{if .State.Health}}{{.State.Health.Status}}{{else}}none{{end}}' \

"$cid" 2>/dev/null || echo unknown)"

case "$status" in

healthy)

return 0

;;

none)

# A service that declares no healthcheck cannot be probed this way, and

# "running" is the strongest signal available for it.

if service_running "$1"; then return 0; fi

;;

esac

i=$((i + 1))

sleep 1

done

return 1

}

# Brings the app back, and reports whether it is actually back.

# Returns 0 only when the app is running AND healthy. APP_STOPPED is cleared at

# that point and not one statement earlier, which is the whole design: the EXIT

# trap calls this too, so leaving the flag set is what makes a failed attempt

# get RETRIED rather than silently abandoned.

# Clearing the flag first — as this did — produced the exact mirror image of the

# bug that put it before `docker compose stop`. A failed `start` cleared the

# flag, warned, and returned 0; the trap then saw 0 and did nothing; and the

# script printed "Backup written to …" and exited 0 with the app down. Measured

# with a shim failing only `compose start lyraflow`:

# Starting the app again...

# WARNING: could not restart the app. …

# Backup written to /tmp/probe-…/2026-08-10T150922Z

# backup.sh EXIT CODE = 0

# lyraflow exited Exited (0) 2 seconds ago

# That is worse than the bug it mirrors, which at least exited non-zero: a cron

# wrapper testing $? saw success while the site went uninstrumented.

# It is reachable by the operator action this script's own comments name. bash

# DEFERS SIGINT while a foreground `docker compose` child is running, so Ctrl-C

# during the drain does not interrupt the stop — it is delivered to a later

# `docker compose` invocation. When that is the restart, the restart dies with

# 130 while every other command in the run succeeded. Reproduced here with a

# logging shim, where the deferred signal landed on a manifest query instead:

# rc=0 compose exec -T postgres … pg_dump …

# rc=130 compose exec -T clickhouse … SELECT count() FROM `events` FINAL

# rc=0 compose start lyraflow

# A manual `docker compose start` seconds later returns 0, so one retry from the

# trap recovers it — which is exactly what leaving the flag set now buys.

start_app_if_stopped() {

[ "$APP_STOPPED" = "1" ] || return 0

# The start comes before ANY write, including the one announcing it. The

# announcement used to be the first statement, and on a closed stdout that

# single `echo` was enough to kill the shell inside the trap before the app

# was ever started -- the fourth consecutive round in which the ordering of

# this function was the defect. `say`/`note` cannot fail either, so neither

# the ordering nor the write-safety is load-bearing on its own.

if ! docker compose start "$APP_SERVICE" >/dev/null 2>&1; then

note "WARNING: 'docker compose start $APP_SERVICE' failed; it will be retried."

return 1

say "Starting the app again..."

# `start` returns as soon as the container is running, which is not the same

# as the app answering requests. Anything an operator runs straight after

# this script — a smoke test, a monitoring probe, the round-trip test — would

# otherwise hit a still-booting server and read it as a failed backup.

# A health timeout is also a failure, not a warning: "the container is up but

# the app never came back" is the same outcome for the site as "the container

# never started", and the caller must not report success for either. The cost

# is that a genuinely slow boot can be waited for twice, once here and once

# from the trap's retry.

if ! wait_until_healthy "$APP_SERVICE" 180; then

note "WARNING: the app was started but has not become healthy."

note "Check: docker compose logs $APP_SERVICE"

return 1

APP_STOPPED=0

}

# Removes the timestamped output directory unless a complete backup landed in

# it.

# Validation creates the directory before the quiesce, and the three artefacts

# appear one at a time, so every failure from validation onwards used to leave

# something behind that reads as a backup to anyone running `ls`. Measured, one

# injected failure per step:

# pg_dump fails -> postgres.dump 0 bytes + clickhouse.zip 8390 bytes

# copy-out fails -> clickhouse.zip 0 bytes

# manifest fails -> three files, only rows.postgres.* missing

# Only the last of those was handled. `restore.sh` refuses a directory with no

# MANIFEST, so none of them is as dangerous as a manifest that is present and

# short — but "a failed run leaves nothing" is either true or it is not, and it

# is far easier to keep true than to document the exceptions.

# Guarded on OUT_CREATED so it can only ever remove a directory this run

# created: validation refuses to start when the directory already exists, and

# creates it with a non-recursive `mkdir` that exactly one racing run can win.

# Cannot fail. Called from the EXIT trap, where a non-zero return under `set -e`

# would abort the trap -- and this used to end in a bare `rm -rf "$OUT"`, which

# on a destination gone read-only aborted cleanup before the app was restarted.

# The trap now runs the restart first and guards every call, so this is the

# third of three independent reasons that cannot happen; it is here so that no

# single one of them is load-bearing.

remove_incomplete_output() {

[ "${OUT_CREATED:-0}" = "1" ] || return 0

if [ "${BACKUP_COMPLETE:-0}" = "1" ]; then return 0; fi

rm -rf "$OUT" 2>/dev/null || {

note "WARNING: could not remove the incomplete backup at $OUT"

note "It has no manifest and cannot be restored; delete it by hand."

}

return 0

}

# The backups disk lives inside the ClickHouse data volume, so the scratch

# archive survives `docker compose down` and accumulates one full backup per

# night until the volume fills. Removing it is load-bearing, not tidiness.

# Removes ONLY an archive this run created, which is why the flag exists rather

# than a bare `rm -f "$CH_FILE"`. The archive name is derived from a UTC

# *second*, so two backups launched 50ms apart compute the same name — and the

# one that loses the BACKUP_ALREADY_EXISTS race would otherwise delete the

# winner's archive on its way out. Measured: run A lost the race and ran the

# unguarded cleanup, leaving B with its three files only because B had already

# copied out; had A's cleanup landed between B's BACKUP and B's copy-out, both

# runs would have failed.

# The cost of the flag is that an archive left behind by a BACKUP that failed

# part-way is not cleaned up. That is the right trade: a stale archive is

# visible, bounded and removable, while deleting a concurrent run's archive

# corrupts a backup that was otherwise going to succeed.

# NOTE FOR restore.sh: this function is shared, and the flag defaults to 0, so

# calling it without setting the flag is a silent no-op. A restore copies an

# archive INTO the container and must therefore set CH_ARTEFACT_CREATED=1

# — otherwise the restored archive stays on the backups disk forever, which is

# precisely the volume growth this exists to prevent.

# BEFORE that copy, not after it, which is the reverse of the rule above and

# for a reason that does not apply there: a restore's archive name comes from

# the manifest's timestamp rather than from the clock, so a second attempt

# recomputes the same name on purpose and overwrites its own leftovers. There

# is no concurrent run whose archive it could take. Setting the flag after the

# copy instead leaks a partial archive from every copy that fails part way —

# measured at 34 kB per abandoned run on a scale model, on a disk that survives

# `docker compose down`.

# Cannot fail, for the same reason as remove_incomplete_output.

remove_in_container_artefact() {

[ "${CH_ARTEFACT_CREATED:-0}" = "1" ] || return 0

[ -n "${CH_FILE:-}" ] || return 0

docker compose exec -T "$CH_SERVICE" \

rm -f "$CH_BACKUP_DIR/$CH_FILE" < /dev/null >/dev/null 2>&1 || true

return 0

}

# ---------------------------------------------------------------------------

# Checksums

# ---------------------------------------------------------------------------

# macOS ships `shasum`, not `sha256sum`, and this script is written for an

# operator on a Mac. Checked up front by backup.sh so a missing hasher is a

# validation failure rather than a failure after the app is already down.

have_sha256() {

command -v sha256sum >/dev/null 2>&1 ||

command -v shasum >/dev/null 2>&1 ||

command -v openssl >/dev/null 2>&1

}

sha256_of() {

if command -v sha256sum >/dev/null 2>&1; then

sha256sum "$1" | awk '{print $1}'

elif command -v shasum >/dev/null 2>&1; then

shasum -a 256 "$1" | awk '{print $1}'

else

openssl dgst -sha256 "$1" | awk '{print $NF}'

}

# ---------------------------------------------------------------------------

# Row counts

# ---------------------------------------------------------------------------

# ch_count_expression <engine>

# MEASURED: a plain `count()` is not stable across BACKUP and RESTORE for a

# merging engine. It is a race, and one that is always eventually lost — the

# collapse is not done by RESTORE, it is done by the background merges that

# follow it. Sampled across three rounds, four phases each:

# BEFORE-RESTORE device_index count=28 FINAL=16 parts=5

# IMMEDIATELY-AFTER device_index count=28 FINAL=16 parts=5 <- unchanged

# +20s device_index count=16 FINAL=16 parts=3 <- the merge

# So a verification run soon after a restore can read the pre-merge number and

# pass, and the same check minutes later reads a smaller one and reports data

# loss. `count()` moved on 6 of 12 samples; `count() … FINAL` was invariant on

# all 12.

# `count() … FINAL` applies the engine's merge logic at read time, so it

# reports the fully-merged count — a property of the data rather than of the

# current part layout. That is the merge-stable expression, recorded per table

# in the manifest as `rowexpr.clickhouse.<table>` so restore.sh compares like

# with like instead of applying a tolerance. It is not free: at 20M rows,

# `FINAL` cost 243ms and 263 MiB read against 1ms for a bare `count()`. That is

# spent inside the quiesce, so it is downtime.

# FINAL is not usable everywhere: `SELECT count() FROM events_dead_letter

# FINAL` is rejected with `Code: 181 … Storage MergeTree doesn't support FINAL

# (ILLEGAL_FINAL)`. A plain MergeTree never collapses rows on merge, so its

# `count()` is stable under merges — but see ch_row_counts on TTL, which is a

# separate way for a plain MergeTree's count to fall. Hence the split below, by

# property rather than by table name: anything ending in MergeTree that is not

# the plain or plain-replicated engine collapses on merge and needs FINAL.

ch_count_expression() {

case "$1" in

MergeTree | ReplicatedMergeTree) echo 'count()' ;;

*MergeTree) echo 'count() FINAL' ;;

*) echo 'count()' ;;

esac

}

# Emits `rows.clickhouse.<table>=<n>`, `rowexpr.clickhouse.<table>=<expr>`, and

# `rowttl.clickhouse.<table>=<expr>` for the tables that carry a table-level TTL.

# The table list comes from system.tables rather than a hardcoded list, because

# a hardcoded list is a list that a future migration silently falls off. The

# three non-data engines are excluded by name and everything else is counted:

# excluding known non-tables fails towards counting a new engine, where

# including only known table engines would fail towards skipping one.

# Materialized views are excluded because they are not storage — every one of

# them has a TO target that is already counted, so counting them would

# double-count the same rows. Dictionaries are excluded because counting one

# forces a load from Postgres, which is neither a property of the ClickHouse

# backup nor something to trigger during a quiesce.

# TTL, and why the answer is to record it rather than to skip the table.

# `events_dead_letter` carries `TTL toDateTime(received_at) + toIntervalDay(30)`.

# Its rows are deleted by the same background merges discussed above, so its

# count falls on its own — measured on a scale model with the TTL shortened:

# manifest 10, count immediately after RESTORE 10, count after the first merge

# 0. Restoring a backup older than the retention window is an ordinary

# disaster-recovery case, and a naive comparison would report catastrophic data

# loss that never happened. FINAL is not available as a fix here (Code 181).

# Dropping the table from the manifest would fix the false alarm and create a

# worse problem: a table missing from `rows.clickhouse.*` is indistinguishable

# from the stale-hardcoded-list failure the enumeration above exists to

# prevent, and the count at backup time is a true fact worth recording.

# So the table is counted like any other and its TTL is recorded alongside.

# The contract for restore.sh: for a table with a `rowttl.clickhouse.*` key, a

# restored count LOWER than the manifest is expected and must not be reported

# as loss; for every other table, equality is required.

# system.tables has no TTL column in 24.8 — the table-level TTL is only

# available inside `engine_full`, which carries the engine clause, PARTITION

# BY, ORDER BY, TTL and SETTINGS and nothing else. Column-level TTLs live in

# the column definitions in `create_table_query` and correctly do not appear

# here, since they expire values rather than rows.

# KNOWN LIMITATION, and one restore.sh must not paper over. This records ANY

# table-level TTL, including `TTL … TO DISK 'x'` and `TO VOLUME 'y'`, which

# MOVE rows between storage tiers rather than deleting them. A table whose only

# TTL is a move would be marked "lower is expected" and have its row

# verification effectively disabled forever, silently. Lyraflow has no such TTL

# today (`events_dead_letter`'s is a plain delete). If one is ever added, this

# must learn to distinguish the two rather than restore.sh loosening further.

ch_row_counts() {

local tables name engine ttl expr suffix count

tables="$(ch_query "

SELECT name, engine,

trim(replaceRegexpOne(extract(engine_full, ' TTL .*'), ' SETTINGS .*\$', ''))

FROM system.tables

WHERE database = '$CH_DATABASE'

AND engine NOT IN ('MaterializedView', 'View', 'Dictionary')

ORDER BY name

")" || return 1

[ -n "$tables" ] || return 1

printf '%s\n' "$tables" | while IFS="$(printf '\t')" read -r name engine ttl; do

[ -n "$name" ] || continue

expr="$(ch_count_expression "$engine")"

suffix=''

case "$expr" in *FINAL) suffix=' FINAL' ;; esac

# The `%s` below are printf placeholders, not shell expansions, and the

# single quotes are what stop the backticks — ClickHouse's identifier

# quoting — from being read as command substitution.

# shellcheck disable=SC2016

count="$(ch_query "$(printf 'SELECT count() FROM `%s`%s' "$name" "$suffix")")" || exit 1

[ -n "$count" ] || exit 1

printf 'rows.clickhouse.%s=%s\n' "$name" "$count"

printf 'rowexpr.clickhouse.%s=%s\n' "$name" "$expr"

# Emitted only for the tables that have one, so its presence is the signal.

if [ -n "$ttl" ]; then

printf 'rowttl.clickhouse.%s=%s\n' "$name" "${ttl#TTL }"

done

}

# Proves the whole Postgres path works before anything is stopped.

# `pg_row_counts` below is the only thing in this file that needs more from

# Postgres than a plain SELECT, and until this existed the first exercise of

# any Postgres query at all happened AFTER the app was down — so a Postgres

# built without libxml, or a psql that cannot authenticate, turned into

# downtime plus a failed backup instead of a refusal. Same principle as

# have_sha256: move the failure to the safe side of the quiesce. Measured at

# ~0.13s, which is noise next to a 30-second drain.

pg_can_count_rows() {

local probe

probe="$(pg_query \

"SELECT (xpath('/row/c/text()', query_to_xml('SELECT 1 AS c', false, true, '')))[1]::text")" ||

return 1

[ "$probe" = "1" ]

}

# Emits `rows.postgres.<table>=<n>`.

# One statement rather than one round trip per table, so every count comes from

# a single snapshot and the numbers in the manifest are mutually consistent.

# `query_to_xml` is the standard way to count a dynamically-named relation from

# plain SQL; it is core Postgres, and the image is pinned by this repo's own

# compose file.

# Views are excluded (`table_type = 'BASE TABLE'`): each identity dictionary's

# source view — identity_bindings_dict_src and friends — is a windowed

# projection of a table already counted here, so counting it would both

# double-count and record a number that is a function of the wall clock.

# Postgres rows are always a plain `count(*)`: nothing in Postgres collapses

# rows on its own, so there is no per-table `rowexpr.postgres.*` key.

pg_row_counts() {

pg_query "

SELECT 'rows.postgres.' || table_name || '=' ||

(xpath('/row/c/text()',

query_to_xml(format('SELECT count(*) AS c FROM public.%I', table_name),

false, true, '')))[1]::text

FROM information_schema.tables

WHERE table_schema = 'public' AND table_type = 'BASE TABLE'

ORDER BY table_name

}

# ---------------------------------------------------------------------------

# Manifest

# ---------------------------------------------------------------------------

# write_manifest <out-dir>

# Plain `key=value` lines, sorted, so `diff` between two manifests reads

# cleanly. Values are never quoted.

# A PARSER MUST SPLIT ON THE FIRST `=` AND TAKE THE WHOLE REMAINDER, not split

# on every `=`. Keys never contain one; values can. `rowexpr.*` values contain

# spaces and parentheses, and a `rowttl.*` value is an arbitrary ClickHouse

# expression — a `TTL … GROUP BY … SET x = max(y)` produces one containing

# spaces, parentheses AND an `=`. Anything that assumes two fields will

# silently truncate it.

# Both checksums are taken from the bytes in <out-dir>, never from the

# in-container archive: restore.sh has to be able to catch an artefact that was

# truncated on its way out, while the live data still exists, and a checksum of

# anything other than the file on disk cannot do that.

write_manifest() {

local out="$1"

local schema_version app_image ch_image pg_image ch_sum pg_sum tmp

tmp="$out/MANIFEST.tmp"

schema_version="$(pg_query 'SELECT max(version) FROM schema_migrations')" || return 1

[ -n "$schema_version" ] || return 1

app_image="$(service_image "$APP_SERVICE")"

ch_image="$(service_image "$CH_SERVICE")"

pg_image="$(service_image "$PG_SERVICE")"

ch_sum="$(sha256_of "$out/clickhouse.zip")" || return 1

pg_sum="$(sha256_of "$out/postgres.dump")" || return 1

if [ -z "$ch_sum" ] || [ -z "$pg_sum" ]; then return 1; fi

# Written to a temporary name and renamed, never straight to MANIFEST. The

# redirect creates its target BEFORE the pipeline runs, and `sort` writes

# whatever it received before the left-hand side died — so a failure part-way

# through leaves a short but perfectly well-formed manifest. Measured: forcing

# a failure inside pg_row_counts produced three files, correct checksums,

# `mode=quiesced` and a matching timestamp, missing only the entire

# `rows.postgres.*` block. That satisfies every assertion a caller is likely

# to make, and a restore.sh iterating those keys would find none and report a

# flawless Postgres verification.

if ! {

printf 'lyraflow_backup_version=1\n'

printf 'timestamp=%s\n' "$STAMP"

printf 'mode=quiesced\n'

printf 'schema_version=%s\n' "$schema_version"

printf 'app_image=%s\n' "$app_image"

printf 'clickhouse_image=%s\n' "$ch_image"

printf 'postgres_image=%s\n' "$pg_image"

printf 'clickhouse_sha256=%s\n' "$ch_sum"

printf 'postgres_sha256=%s\n' "$pg_sum"

ch_row_counts || exit 1

pg_row_counts || exit 1

} | LC_ALL=C sort | artefact_write "$tmp"; then

rm -f "$tmp"

return 1

# `mv`, not a second write: the rename carries the mode artefact_write gave

# the temporary file, so MANIFEST cannot appear with any other mode, and it

# appears atomically or not at all.

mv "$tmp" "$out/MANIFEST"

}

FilesExpand file tree

backup-lib.sh

Latest commit

History

backup-lib.sh

File metadata and controls