трек 2: read-only аудит инвариантов I1-I7 и автоматическая матрица отказов

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

audit: БД только READ ONLY-транзакциями, S3 ключом auditor (Read+List): потерянная работа,
один эффект на версию, опубликованные объекты и SHA-256, монотонная версия, источник у
активной работы, посторонние объекты, попытки. pipeline_faults: дубль, reorder, crash relay
после confirm, crash worker до/после побочного эффекта и после COMMIT, временные ошибки,
два worker, устаревшая версия (v2 медленная, v3 быстрая; дубль v1), reorder версий в очереди,
зависший worker (docker pause).
This commit is contained in:
Alexey Bodanyuk
2026-10-31 17:17:16 +03:00
parent a3fa7713fa
commit e453ee2312
2 changed files with 499 additions and 0 deletions
+309
View File
@@ -0,0 +1,309 @@
#!/usr/bin/env python3
"""Track 2 experiment: proven idempotency of the versioned pipeline under injected faults.
Hypothesis. With at-least-once delivery, duplicated and reordered messages, crashes of the worker
before/after its side effects and stale jobs of an older pipeline version:
H1 no accepted (202) work is lost - every upload ends ready (or failed with a code, never stuck);
H2 every effect (attachment, pipeline_version, variant) is published exactly once - one set of
variant rows written by ONE successful attempt, one set of objects in the bucket;
H3 the published version only grows - an older job never overwrites a newer result, `ready`
never goes back to queued/processing;
H4 objects of interrupted attempts are garbage that the cleanup removes; nothing published is
ever deleted.
Method. For every scenario the affected service (relay/worker) is recreated with FAULT_POINTS of
the scenario (the real code paths, real crashes = os._exit of the process, real `docker pause`),
N uploads go through the LB, then the scenario-specific facts are read from PostgreSQL and S3 and
the read-only invariant audit (messenger.audit, I1-I7, with SHA-256 of every published object)
runs over the WHOLE stand. The normal configuration is restored after each scenario.
python bench/pipeline_faults.py --output evidence/track2 [--uploads 5]
"""
from __future__ import annotations
import argparse
import datetime as dt
import json
import platform
import statistics
import subprocess
import sys
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "student-tests"))
import imagegen # noqa: E402
from conftest import (Api, attachment_row, compose, events_of, jobctl, psql, s3_keys, service_with_env, # noqa: E402
wait_healthy)
FAST = {"JOB_LEASE_SECONDS": "6", "ORPHAN_GRACE_SECONDS": "3", "CLEANUP_INTERVAL_SECONDS": "1"}
def now() -> str:
return dt.datetime.now(dt.timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
def wait(condition, timeout: float, what: str):
deadline = time.monotonic() + timeout
while True:
value = condition()
if value:
return value
if time.monotonic() > deadline:
raise AssertionError(what)
time.sleep(0.5)
def facts(attachment_id: str, objects: set[str]) -> dict:
row = attachment_row(attachment_id)
out = psql("SELECT json_build_object("
"'claims', (SELECT count(*) FROM job_attempts a JOIN jobs j ON j.id = a.job_id WHERE j.attachment_id = '{0}'), "
"'succeeded_attempts', (SELECT count(*) FROM job_attempts a JOIN jobs j ON j.id = a.job_id "
" WHERE j.attachment_id = '{0}' AND a.outcome = 'succeeded'), "
"'variant_rows', (SELECT count(*) FROM attachment_variants WHERE attachment_id = '{0}'), "
"'versions', (SELECT coalesce(json_agg(DISTINCT pipeline_version), '[]') FROM attachment_variants "
" WHERE attachment_id = '{0}'), "
"'job_states', (SELECT coalesce(json_agg(state || ':v' || pipeline_version ORDER BY created_at), '[]') "
" FROM jobs WHERE attachment_id = '{0}'), "
"'uncleaned_failed_attempts', (SELECT count(*) FROM job_attempts a JOIN jobs j ON j.id = a.job_id "
" WHERE j.attachment_id = '{0}' AND a.outcome <> 'succeeded' AND a.objects_cleaned_at IS NULL))"
.format(attachment_id))
f = json.loads(out)
f.update(status=row["status"], published_version=row["published_version"], staging_state=row["staging_state"],
objects=len([k for k in objects if k.startswith(f"a/{attachment_id}/")]))
f["events"] = [e["event"] for e in events_of(attachment_id)]
return f
def audit() -> dict:
r = compose("run", "--rm", "--no-deps", "-T", "auditor", "--verify-checksums", check=False, timeout=600)
report = json.loads(r.stdout)
return {"ok": report["ok"], "violations": report["violations"], "notes": len(report["notes"]),
"objects": report["objects"]}
class Run:
def __init__(self, api: Api, uploads: int):
self.api = api
self.n = uploads
self.alice, self.bob = api.user("t2a"), api.user("t2b")
self.chat = api.chat(self.alice, self.bob)
def upload_many(self, n: int | None = None) -> list[dict]:
return [self.api.upload(self.alice, self.chat, imagegen.png(60 + i, 40, noise=True)).json()
for i in range(n or self.n)]
def finish(self, atts: list[dict], timeout: float = 180) -> tuple[list[dict], list[float]]:
started = time.monotonic()
took = []
for a in atts:
self.api.wait_status(self.bob, a["id"], timeout=timeout)
took.append(round(time.monotonic() - started, 2))
return atts, took
def expect_one_effect(f: dict, version: int = 1) -> list[str]:
"""H1 + H2 for a plain upload of pipeline `version`."""
problems = []
if f["status"] != "ready":
problems.append(f"status {f['status']}")
if f["succeeded_attempts"] != 1 or f["variant_rows"] != 3 or f["published_version"] != version:
problems.append(f"effect {f['succeeded_attempts']} attempts/{f['variant_rows']} rows/v{f['published_version']}")
if f["objects"] != 4: # original + thumbnail + crop + optimized, cleanup done
problems.append(f"{f['objects']} objects in the bucket")
return problems
def plain_scenario(run: Run, name: str, service: str, env: dict, *, scale: int | None = None,
cleanup_wait: bool = True) -> dict:
with service_with_env(service, scale=scale, **env):
atts, took = run.finish(run.upload_many())
if cleanup_wait: # garbage of interrupted attempts: cleaned after ORPHAN_GRACE_SECONDS
try:
wait(lambda: all(json.loads(psql(
"SELECT json_build_object('n', count(*)) FROM job_attempts a JOIN jobs j ON j.id = a.job_id "
f"WHERE j.attachment_id = '{a['id']}' AND a.outcome <> 'succeeded' "
"AND a.objects_cleaned_at IS NULL"))["n"] == 0 for a in atts), 60, "cleanup")
except AssertionError:
pass # reported below as uncleaned attempts
objects = s3_keys("a/")
per = [facts(a["id"], objects) for a in atts]
problems = {a["id"]: p for a, f in zip(atts, per) if (p := expect_one_effect(f))}
return {"scenario": name, "faults": {service: env}, "uploads": len(atts),
"ready": sum(f["status"] == "ready" for f in per), "claims_total": sum(f["claims"] for f in per),
"max_claims": max(f["claims"] for f in per), "time_to_ready_s": {"p50": statistics.median(took),
"max": max(took)},
"event_sequences": sorted({" > ".join(f["events"]) for f in per}), "problems": problems}
def versions_scenario(run: Run) -> dict:
"""Stale job after a newer version: v2 (slow) and v3 run at the same time; v3 publishes first,
v2 finishes later and must end superseded without touching the published v3 result."""
att = run.upload_many(1)[0]
run.api.wait_status(run.bob, att["id"], "ready")
observed = set()
with service_with_env("worker", WORKER_FAULT_POINTS="slow=10,slow_pipeline=2"):
jobctl("reprocess", att["id"], "--pipeline", "2", "--operator", "track2", "--reason", "new encoder")
wait(lambda: "processing:v2" in facts(att["id"], set())["job_states"], 60, "v2 not claimed")
jobctl("reprocess", att["id"], "--pipeline", "3", "--operator", "track2", "--reason", "newer encoder")
deadline = time.monotonic() + 120
while time.monotonic() < deadline:
meta = run.api.attachment(run.bob, att["id"])
observed.add(meta["status"])
states = facts(att["id"], set())["job_states"]
if not any(s.startswith(("queued", "processing")) for s in states):
break
time.sleep(0.3)
final = run.api.attachment(run.bob, att["id"])
f = facts(att["id"], s3_keys("a/"))
thumb_v3 = psql(f"SELECT encode(sha256, 'hex') FROM attachment_variants WHERE attachment_id = '{att['id']}' "
"AND pipeline_version = 3 AND name = 'thumbnail'")
problems = []
if f["job_states"] != ["succeeded:v1", "superseded:v2", "succeeded:v3"]:
problems.append(f"job states {f['job_states']}")
if f["published_version"] != 3 or observed != {"ready"}:
problems.append(f"published v{f['published_version']}, statuses seen {sorted(observed)}")
if next(v["sha256"] for v in final["variants"] if v["name"] == "thumbnail") != thumb_v3:
problems.append("published thumbnail is not the v3 one")
# Stale duplicate of the FIRST job, published by hand with the relay's broker user: a no-op.
job_v1 = psql(f"SELECT id FROM jobs WHERE attachment_id = '{att['id']}' AND pipeline_version = 1")
script = ("import asyncio, aio_pika, datetime as d\n"
"from messenger import broker\n"
"async def m():\n"
" c = await aio_pika.connect(open('/run/secrets/broker_url').read().strip())\n"
" ch = await c.channel(publisher_confirms=True)\n"
" ex = await ch.get_exchange(broker.EXCHANGE, ensure=False)\n"
f" body = broker.encode('{job_v1}', '{att['id']}', 1, 1, d.datetime.now(d.timezone.utc))\n"
" await ex.publish(aio_pika.Message(body), routing_key=broker.ROUTING_KEY)\n"
" await c.close()\n"
"asyncio.run(m())\n")
compose("exec", "-T", "relay", "python", "-c", script)
time.sleep(3)
after = run.api.attachment(run.bob, att["id"])
if after != final:
problems.append("stale v1 message changed the attachment")
return {"scenario": "stale_version_and_reorder", "uploads": 1, "statuses_seen_during_reprocess": sorted(observed),
"job_states": f["job_states"], "published_version": f["published_version"],
"variant_rows_by_version": f["variant_rows"], "objects": f["objects"],
"events": f["events"], "problems": problems}
def queued_reorder_scenario(run: Run) -> dict:
"""Both newer versions wait in the outbox; the relay publishes them in REVERSE order and one worker
takes them one by one: v3 is published first, v2 is then superseded at its claim (no work)."""
att = run.upload_many(1)[0]
run.api.wait_status(run.bob, att["id"], "ready")
# The worker stays stopped until it is recreated with concurrency 1 (no normal start in between).
# Worker and relay stay stopped until each is recreated with the scenario's settings: no normal
# start in between that would consume/publish the jobs in the ordinary order.
compose("stop", "worker", "relay")
try:
jobctl("reprocess", att["id"], "--pipeline", "2", "--operator", "track2", "--reason", "v2")
jobctl("reprocess", att["id"], "--pipeline", "3", "--operator", "track2", "--reason", "v3")
with service_with_env("relay", RELAY_FAULT_POINTS="reverse_order"):
wait(lambda: psql("SELECT count(*) FROM outbox o JOIN jobs j ON j.id = o.job_id WHERE j.attachment_id = "
f"'{att['id']}' AND j.pipeline_version > 1 AND o.published_at IS NOT NULL") == "2",
30, "both reprocess jobs not published") # in one relay batch: v3 first, then v2
with service_with_env("worker", WORKER_CONCURRENCY="1"):
wait(lambda: not any(s.startswith(("queued", "processing"))
for s in facts(att["id"], set())["job_states"]), 120, "reprocess jobs did not finish")
finally:
compose("start", "worker", "relay")
wait_healthy("worker", "relay")
f = facts(att["id"], s3_keys("a/"))
problems = []
if f["job_states"] != ["succeeded:v1", "superseded:v2", "succeeded:v3"] or f["published_version"] != 3:
problems.append(f"{f['job_states']} published v{f['published_version']}")
if f["claims"] != 2: # v1 and v3 only: v2 was superseded before any work
problems.append(f"{f['claims']} claims")
return {"scenario": "reordered_versions_in_queue", "job_states": f["job_states"], "claims": f["claims"],
"published_version": f["published_version"], "events": f["events"], "problems": problems}
def hung_worker_scenario(run: Run) -> dict:
with service_with_env("worker", scale=2, WORKER_FAULT_POINTS="slow=8", **FAST):
att = run.upload_many(1)[0]
owner = wait(lambda: psql(f"SELECT lease_owner FROM jobs WHERE attachment_id = '{att['id']}' "
"AND state = 'processing'"), 30, "not claimed")
ids = compose("ps", "-q", "worker").stdout.split()
names = dict(line.split() for line in subprocess.run(
["docker", "inspect", "-f", "{{.Config.Hostname}} {{.Id}}", *ids], capture_output=True, text=True,
check=True).stdout.splitlines())
subprocess.run(["docker", "pause", names[owner]], check=True)
try:
run.api.wait_status(run.bob, att["id"], "ready", timeout=150)
published = psql(f"SELECT string_agg(object_key, ',' ORDER BY object_key) FROM attachment_variants "
f"WHERE attachment_id = '{att['id']}'")
finally:
subprocess.run(["docker", "unpause", names[owner]], check=True)
time.sleep(15)
again = psql(f"SELECT string_agg(object_key, ',' ORDER BY object_key) FROM attachment_variants "
f"WHERE attachment_id = '{att['id']}'")
wait(lambda: json.loads(psql("SELECT json_build_object('n', count(*)) FROM job_attempts a JOIN jobs j "
f"ON j.id = a.job_id WHERE j.attachment_id = '{att['id']}' AND a.outcome <> "
"'succeeded' AND a.objects_cleaned_at IS NULL"))["n"] == 0, 60, "cleanup")
f = facts(att["id"], s3_keys("a/"))
problems = expect_one_effect(f)
if published != again:
problems.append("the resumed (fenced) attempt changed the published result")
return {"scenario": "hung_worker_docker_pause", "frozen": owner, "claims": f["claims"], "events": f["events"],
"problems": problems}
def main() -> int:
p = argparse.ArgumentParser()
p.add_argument("--output", default="evidence/track2")
p.add_argument("--uploads", type=int, default=5)
args = p.parse_args()
out = ROOT / args.output
out.mkdir(parents=True, exist_ok=True)
api = Api()
run = Run(api, args.uploads)
started = now()
results = []
plan = [
("baseline_no_faults", "worker", {}),
("duplicate_publish", "relay", {"RELAY_FAULT_POINTS": "duplicate_publish"}),
("duplicate_and_reverse_order", "relay", {"RELAY_FAULT_POINTS": "duplicate_publish,reverse_order"}),
("relay_crash_after_confirm_before_mark", "relay", {"RELAY_FAULT_POINTS": "crash_after_publish"}),
("worker_crash_before_side_effect", "worker", {"WORKER_FAULT_POINTS": "crash_after_claim", **FAST}),
("worker_crash_after_s3_before_commit", "worker", {"WORKER_FAULT_POINTS": "crash_after_s3_write", **FAST}),
("worker_crash_after_commit_before_ack", "worker", {"WORKER_FAULT_POINTS": "crash_after_commit", **FAST}),
("transient_errors_then_success", "worker", {"WORKER_FAULT_POINTS": "transient_error=2",
"JOB_RETRY_BASE_SECONDS": "0.5", **FAST}),
]
for name, service, env in plan:
print(f"[{now()}] {name}", flush=True)
results.append(plain_scenario(run, name, service, env))
print(f"[{now()}] two_workers_duplicate_publish", flush=True)
with service_with_env("relay", RELAY_FAULT_POINTS="duplicate_publish"):
results.append(plain_scenario(run, "two_workers_duplicate_publish", "worker", {}, scale=2))
for scenario in (versions_scenario, queued_reorder_scenario, hung_worker_scenario):
print(f"[{now()}] {scenario.__name__}", flush=True)
results.append(scenario(run))
wait_healthy("worker", "relay")
report = audit()
summary = {
"experiment": "track 2: idempotency of the versioned pipeline under injected faults",
"started": started, "finished": now(),
"host": {"os": platform.platform(), "python": platform.python_version()},
"uploads_per_plain_scenario": args.uploads,
"scenarios": results,
"audit_after_all": report,
"violations": sum(len(r["problems"]) for r in results) + len(report["violations"]),
}
(out / "summary.json").write_text(json.dumps(summary, ensure_ascii=False, indent=1), encoding="utf-8")
lines = [f"{'scenario':45} {'uploads':>7} {'ready':>5} {'claims':>6} problems"]
for r in results:
lines.append(f"{r['scenario']:45} {r.get('uploads', 1):>7} {r.get('ready', '-'):>5} "
f"{r.get('claims_total', r.get('claims', '-')):>6} {len(r['problems'])}")
lines.append(f"audit after all scenarios: ok={report['ok']} violations={len(report['violations'])} "
f"objects={report['objects']}")
(out / "summary.txt").write_text("\n".join(lines) + "\n", encoding="utf-8")
print("\n".join(lines))
return 0 if summary["violations"] == 0 else 1
if __name__ == "__main__":
sys.exit(main())
+190
View File
@@ -0,0 +1,190 @@
"""Track 2: read-only checker of the pipeline invariants: `python -m messenger.audit`
(compose service `auditor`, profile tools; `make audit`).
Read-only by construction: every database transaction is `READ ONLY` (PostgreSQL refuses any
write), and the S3 key pair of the auditor has Read + List only. Output: one JSON report; exit 1
if any invariant is violated.
Invariants (docs/PIPELINE.md):
I1 no lost work - a queued/processing attachment has an active job; an active job has an
outbox row (published or due) - nothing waits for a message that nobody sends.
I2 one effect - per (attachment, pipeline_version) either no variant rows or exactly the
three variants, all written by ONE successful attempt of a job of that version.
I3 published data - every published object (original, variants) exists with the stored size
(--verify-checksums: and the stored SHA-256).
I4 monotonic - published_version = the highest version with a succeeded job; a ready
attachment is never queued/processing again (published_version stays set).
I5 source kept - an active job whose original is not published still has its staging copy
(state present AND the object exists in the staging bucket).
I6 no stray data - every object in the buckets is referenced: published, staging of a live row,
or planned by an attempt whose objects were not cleaned yet. An object of an
already cleaned attempt is reported as `leaked_object` (written by a stalled
attempt after its cleanup: garbage, never a loss).
I7 attempts - a succeeded job has exactly one succeeded attempt; at most one running attempt
per job, and it is the lease holder.
"""
from __future__ import annotations
import argparse
import asyncio
import hashlib
import json
import sys
from .config import ConfigError, Settings
from .db import Database
from .errors import Unavailable
from .pipeline import VARIANT_NAMES
from .storage import ObjectMissing, SENTINEL_KEY, Storage
async def rows(db: Database, sql: str, params=()) -> list[tuple]:
async def work(conn):
await conn.execute("SET TRANSACTION READ ONLY")
cur = await conn.execute(sql, params)
return await cur.fetchall()
return await db.run(work, deadline=db.deadline)
async def list_all(storage: Storage, prefix: str) -> dict[str, int]:
found, after = {}, None
while True:
page, more = await storage.list_page(prefix, after, 1000)
for item in page:
found[item.key] = item.size
if not more or not page:
return found
after = page[-1].key
async def audit(settings: Settings, verify: bool) -> dict:
db = Database(settings, name="messenger-audit")
final = Storage(settings)
staging = Storage(settings, bucket=settings.s3_staging_bucket)
await db.open()
await final.open()
await staging.open()
violations: list[dict] = []
notes: list[dict] = []
def bad(kind: str, **detail) -> None:
violations.append({"invariant": kind, **{k: str(v) for k, v in detail.items()}})
try:
# I1: queued/processing attachment without an active job; active job without outbox row.
for (aid,) in await rows(db, "SELECT a.id FROM attachments a WHERE a.status IN ('queued', 'processing') "
"AND NOT EXISTS (SELECT 1 FROM jobs j WHERE j.attachment_id = a.id "
"AND j.state IN ('queued', 'processing'))"):
bad("I1_no_active_job", attachment_id=aid)
for (jid,) in await rows(db, "SELECT j.id FROM jobs j WHERE j.state = 'queued' AND NOT EXISTS "
"(SELECT 1 FROM outbox o WHERE o.job_id = j.id)"):
bad("I1_queued_without_outbox", job_id=jid)
# I2: exactly three variants per published version, one successful attempt of that version.
for aid, version, names, tokens in await rows(
db, "SELECT attachment_id, pipeline_version, array_agg(name ORDER BY name), "
"array_agg(DISTINCT attempt_token) FROM attachment_variants GROUP BY 1, 2"):
if sorted(names) != sorted(VARIANT_NAMES):
bad("I2_variant_set", attachment_id=aid, version=version, names=names)
if len(tokens) != 1:
bad("I2_mixed_attempts", attachment_id=aid, version=version, attempts=len(tokens))
for aid, version, outcome, jversion in await rows(
db, "SELECT v.attachment_id, v.pipeline_version, ja.outcome, j.pipeline_version "
"FROM attachment_variants v JOIN job_attempts ja ON ja.token = v.attempt_token "
"JOIN jobs j ON j.id = ja.job_id WHERE ja.outcome <> 'succeeded' OR j.pipeline_version "
"<> v.pipeline_version OR j.attachment_id <> v.attachment_id"):
bad("I2_variant_from_unsuccessful_attempt", attachment_id=aid, version=version, outcome=outcome)
# I4: published_version = highest succeeded version; ready <=> published.
for aid, status, published, best in await rows(
db, "SELECT a.id, a.status, a.published_version, (SELECT max(j.pipeline_version) FROM jobs j "
"WHERE j.attachment_id = a.id AND j.state = 'succeeded') FROM attachments a "
"WHERE a.published_version IS DISTINCT FROM (SELECT max(j.pipeline_version) FROM jobs j "
"WHERE j.attachment_id = a.id AND j.state = 'succeeded') "
"OR (a.published_version IS NOT NULL AND a.status <> 'ready')"):
bad("I4_published_version", attachment_id=aid, status=status, published=published, best_succeeded=best)
# I7: attempts.
for (jid,) in await rows(db, "SELECT j.id FROM jobs j WHERE j.state = 'succeeded' AND (SELECT count(*) "
"FROM job_attempts a WHERE a.job_id = j.id AND a.outcome = 'succeeded') <> 1"):
bad("I7_succeeded_attempts", job_id=jid)
for jid, token in await rows(db, "SELECT a.job_id, a.token FROM job_attempts a JOIN jobs j ON j.id = a.job_id "
"WHERE a.outcome = 'running' AND j.lease_token IS DISTINCT FROM a.token"):
bad("I7_running_not_holder", job_id=jid, token=token)
# I3 + I5 + I6: storage.
published = {k: (s, h) for k, s, h in await rows(
db, "SELECT object_key, size_bytes, sha256 FROM attachment_variants UNION ALL "
"SELECT object_key, size_bytes, sha256 FROM attachments WHERE object_key IS NOT NULL")}
staging_live = {k: (s, st, status) for k, s, st, status in await rows(
db, "SELECT staging_key, size_bytes, staging_state, status FROM attachments WHERE staging_key IS NOT NULL")}
planned_open, planned_cleaned = set(), set()
for keys, cleaned in await rows(db, "SELECT planned_keys, objects_cleaned_at IS NOT NULL FROM job_attempts"):
(planned_cleaned if cleaned else planned_open).update(keys)
final_objects = await list_all(final, "")
staging_objects = await list_all(staging, "")
for key, (size, sha) in published.items():
if key not in final_objects:
bad("I3_published_missing", key=key)
elif final_objects[key] != size:
bad("I3_published_size", key=key, stored=size, actual=final_objects[key])
elif verify:
data = await final.read_all(key, limit=settings.max_upload_bytes)
if hashlib.sha256(data).digest() != bytes(sha):
bad("I3_published_checksum", key=key)
for key, (size, state, status) in staging_live.items():
# 'pending' may lack its object (the PUT failed: the janitor removes the row); accepted
# work (queued/processing) must have it.
active = status in ("queued", "processing")
if state == "present" and key not in staging_objects and active:
bad("I5_staging_missing", key=key, status=status)
for (aid,) in await rows(db, "SELECT a.id FROM attachments a WHERE a.status IN ('queued', 'processing') "
"AND a.object_key IS NULL AND a.staging_state <> 'present'"):
bad("I5_active_without_source", attachment_id=aid)
for key in final_objects:
if key == SENTINEL_KEY or key in published or key in planned_open:
continue
if key in planned_cleaned:
notes.append({"kind": "leaked_object", "key": key})
else:
bad("I6_unknown_final_object", key=key)
for key in staging_objects:
if key == SENTINEL_KEY:
continue
row = staging_live.get(key)
if row is None:
bad("I6_unknown_staging_object", key=key)
elif row[1] == "deleted":
notes.append({"kind": "staging_marked_deleted_still_present", "key": key})
counts = dict(await rows(db, "SELECT status, count(*) FROM attachments GROUP BY status"))
jobs = dict(await rows(db, "SELECT state, count(*) FROM jobs GROUP BY state"))
return {"ok": not violations, "violations": violations, "notes": notes,
"attachments": counts, "jobs": jobs,
"objects": {"final": len(final_objects), "staging": len(staging_objects),
"published": len(published), "checksums_verified": verify}}
finally:
await staging.close()
await final.close()
await db.close()
def main(argv: list[str] | None = None) -> int:
p = argparse.ArgumentParser(prog="audit", description="Read-only check of the pipeline invariants")
p.add_argument("--verify-checksums", action="store_true", help="read every published object and compare SHA-256")
args = p.parse_args(argv)
try:
settings = Settings.from_env()
except ConfigError as exc:
print(f"configuration error: {exc}", file=sys.stderr)
return 2
try:
report = asyncio.run(audit(settings, args.verify_checksums))
except (Unavailable, ObjectMissing) as exc:
print(json.dumps({"ok": False, "error": "store_unavailable", "detail": type(exc).__name__}))
return 1
print(json.dumps(report, ensure_ascii=False, indent=1, default=str))
return 0 if report["ok"] else 1
if __name__ == "__main__":
sys.exit(main())