lab5 API: upload 202 queued через staging + задание + outbox одной транзакцией

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

Порядок: pending-намерение -> PUT в приватный staging bucket -> COMMIT queued + jobs + outbox
-> 202. Проверки 413/415 lab4 остаются синхронными. GET видит queued/processing/failed,
content: 409 до ready, варианты thumbnail/crop/optimized; к сообщению можно прикрепить
queued/processing, failed -> 409. API не связан с брокером.
This commit is contained in:
Alexey Bodanyuk
2026-10-31 16:12:52 +03:00
parent a633949919
commit 655f264104
6 changed files with 176 additions and 104 deletions
+59 -34
View File
@@ -22,7 +22,7 @@ from starlette.types import ASGIApp, Message as AsgiMessage, Receive, Scope, Sen
from . import attachments, authhdr, cursor, docs, passwords, repo, validation
from .config import Settings
from .db import Database
from .errors import BASIC_CHALLENGE, ApiError, InvalidRequest, NotFound, PayloadTooLarge, Unauthorized, Unavailable
from .errors import BASIC_CHALLENGE, ApiError, Conflict, InvalidRequest, NotFound, PayloadTooLarge, Unauthorized, Unavailable
from .sessions import SessionStore, build_store
from .storage import ObjectMissing, Storage
from .timefmt import rfc3339
@@ -132,6 +132,10 @@ def storage_of(request: Request) -> Storage:
return request.app.state.storage
def staging_of(request: Request) -> Storage:
return request.app.state.staging
async def authenticate(request: Request) -> str:
"""User id of a valid Bearer session, otherwise 401. X-User-Id is ignored since lab 2.
@@ -363,16 +367,17 @@ async def list_messages(request: Request) -> JSONResponse:
return ok({"items": [m.to_json() for m in items], "next_cursor": next_cursor})
# ------------------------------------------------------------------ attachments (lab 4)
# ------------------------------------------------------------------ attachments (lab 4, lab 5)
def fault(request: Request, point: str) -> None:
"""Failure injection for the level 4 experiments (FAULT_POINTS, compose.faults.yaml only)."""
"""Failure injection for the experiments (FAULT_POINTS of the api service, compose variable
API_FAULT_POINTS; empty in normal runs)."""
if point in request.app.state.settings.fault_points:
log.warning("fault_injected", extra={"fields": {"request_id": request.state.request_id, "point": point}})
raise Unavailable("Внедрённый отказ: " + point)
async def upload_attachment(request: Request) -> JSONResponse:
"""POST /chats/{chat_id}/attachments?filename=...: raw PNG/JPEG -> 201 ready (attachments.py)."""
"""POST /chats/{chat_id}/attachments?filename=...: raw PNG/JPEG -> 202 queued (attachments.py)."""
settings: Settings = request.app.state.settings
reader = attachments.BodyReader(request)
slots: asyncio.Semaphore = request.app.state.upload_slots
@@ -406,23 +411,25 @@ async def upload_attachment(request: Request) -> JSONResponse:
async def store_upload(request: Request, db: Database, chat_id: str, user_id: str, filename: str,
upload: attachments.Upload) -> JSONResponse:
"""intent (pending) -> S3 PUT -> ready -> 201 (order and failure analysis: docs/CONSISTENCY.md)."""
"""intent (pending) -> staging PUT -> queued + job + outbox -> 202 (docs/PIPELINE.md).
No resize here: the worker does it; a stopped worker only delays `ready`."""
settings: Settings = request.app.state.settings
attachment_id = uuid.uuid4()
object_key = attachments.new_object_key()
storage = storage_of(request)
staging_key = attachments.new_staging_key()
await attachments.create_intent(db, attachment_id, chat_id, user_id, filename, upload.info, len(upload.data),
upload.sha256, object_key)
await storage.put(object_key, upload.data, upload.info.content_type,
{"attachment-id": str(attachment_id), "sha256": upload.sha256.hex()})
upload.sha256, staging_key)
await staging_of(request).put(staging_key, upload.data, upload.info.content_type,
{"attachment-id": str(attachment_id), "sha256": upload.sha256.hex()})
fault(request, "after_s3_put")
attachment = await attachments.mark_ready(db, attachment_id)
attachment = await attachments.enqueue(db, attachment_id, settings)
if attachment is None:
raise Unavailable("Загрузка не завершена, повторите запрос")
fault(request, "after_ready_commit")
log.info("attachment_created", extra={"fields": {"request_id": request.state.request_id,
"attachment_id": attachment.id, "chat_id": chat_id,
"size": attachment.size_bytes, "type": attachment.content_type}})
return ok(attachment.to_json(), 201)
fault(request, "after_enqueue_commit")
log.info("attachment_queued", extra={"fields": {"request_id": request.state.request_id,
"attachment_id": attachment.id, "chat_id": chat_id,
"size": attachment.size_bytes, "type": attachment.content_type,
"pipeline_version": settings.pipeline_version}})
return ok(attachment.to_json(), 202)
async def get_attachment(request: Request) -> JSONResponse:
@@ -439,46 +446,60 @@ def content_disposition(filename: str) -> str:
async def download_attachment(request: Request) -> Response:
"""Original bytes through the authorized API: no redirect, no public S3 URL. Streamed from
S3 in 64 KiB chunks (memory per download does not depend on the file size)."""
"""Original or variant bytes through the authorized API: no redirect, no public S3 URL.
queued/processing/failed -> 409 (not ready); a variant a ready attachment lacks -> 404.
Streamed from S3 in 64 KiB chunks (memory per download does not depend on the file size)."""
user_id = await authenticate(request)
attachment_id = path_uuid(request, "attachment_id")
query = validation.query_items(request.scope["query_string"])
try:
validation.variant_param(query)
except validation.NotFoundVariant:
variant = validation.variant_param(query)
except ApiError:
await attachments.get_visible(db_of(request), user_id, attachment_id) # 404 for outsiders first
raise NotFound("Такого варианта у вложения нет") from None
raise
attachment = await attachments.get_visible(db_of(request), user_id, attachment_id)
if attachment.status != "ready":
raise Conflict("Обработка вложения не удалась" if attachment.status == "failed" else "Вложение ещё не готово")
if attachment.lost:
raise NotFound("Оригинал вложения утрачен")
if variant == "original":
key, size_expected, media_type, etag = (attachment.object_key, attachment.size_bytes,
attachment.content_type, attachment.sha256)
name = attachment.filename
else:
found = attachment.variant(variant)
if found is None:
raise NotFound("Такого варианта у вложения нет")
key, size_expected, media_type, etag = found.object_key, found.size_bytes, found.content_type, found.sha256
stem = attachment.filename.rsplit(".", 1)[0] or "image"
name = stem + "-" + variant + (".webp" if media_type == "image/webp" else ".jpg")
try:
size, chunks = await storage_of(request).open_stream(attachment.object_key)
size, chunks = await storage_of(request).open_stream(key)
except ObjectMissing:
# A ready row without its object: data loss, reported by the reconciler (quarantine).
log.error("attachment_object_missing", extra={"fields": {"request_id": request.state.request_id,
"attachment_id": attachment.id}})
raise NotFound("Оригинал вложения утрачен") from None
if size != attachment.size_bytes:
"attachment_id": attachment.id, "variant": variant}})
raise Unavailable("Файл вложения временно недоступен") from None
if size != size_expected:
log.error("attachment_size_mismatch", extra={"fields": {"request_id": request.state.request_id,
"attachment_id": attachment.id}})
"attachment_id": attachment.id, "variant": variant}})
await chunks.aclose()
raise Unavailable("Оригинал вложения временно недоступен")
raise Unavailable("Файл вложения временно недоступен")
headers = {
"Content-Length": str(size),
"Content-Disposition": content_disposition(attachment.filename),
"Content-Disposition": content_disposition(name),
"X-Content-Type-Options": "nosniff",
"Cache-Control": "private, no-store",
"ETag": '"' + attachment.sha256 + '"',
"ETag": '"' + etag + '"',
}
return StreamingResponse(chunks, media_type=attachment.content_type, headers=headers)
return StreamingResponse(chunks, media_type=media_type, headers=headers)
def create_app(settings: Settings | None = None) -> Starlette:
settings = settings or Settings.from_env()
db = Database(settings)
sessions = build_store(settings, db)
storage = Storage(settings)
storage = Storage(settings) # final bucket: Read
staging = Storage(settings, bucket=settings.s3_staging_bucket) # staging bucket: Write (+ janitor DELETE)
async def purge_loop() -> None:
"""Housekeeping: expired Idempotency-Key rows (older than IDEMPOTENCY_TTL_SECONDS)."""
@@ -492,12 +513,12 @@ def create_app(settings: Settings | None = None) -> Starlette:
pass # database unavailable right now; next round
async def janitor_loop(every: float) -> None:
"""Level 4 cleanup: intents of failed uploads older than PENDING_GRACE_SECONDS (attachments.py).
"""Cleanup of intents of failed uploads older than PENDING_GRACE_SECONDS (attachments.py).
Every replica runs it; the advisory lock lets one round run at a time."""
while True:
await asyncio.sleep(every * (0.5 + random.random()))
try:
removed = await attachments.clean_stale_pending(db, storage, settings.pending_grace_seconds)
removed = await attachments.clean_stale_pending(db, staging, settings.pending_grace_seconds)
if removed:
log.info("stale_uploads_removed", extra={"fields": {"count": removed}})
except ApiError:
@@ -524,11 +545,13 @@ def create_app(settings: Settings | None = None) -> Starlette:
passwords.configure(settings.hash_concurrency)
await db.open()
await storage.open()
await staging.open()
log.info("dependencies_configured", extra={"fields": {
"session_store": sessions.name, "session_ttl_s": settings.session_ttl_seconds,
"db_pool_max": settings.db_pool_max, "db_timeout_s": settings.db_timeout_seconds,
"instance_id": settings.instance_id, "drain_s": settings.drain_seconds,
"message_write_path": settings.message_write_path, "s3_bucket": settings.s3_bucket,
"s3_staging_bucket": settings.s3_staging_bucket, "pipeline_version": settings.pipeline_version,
"upload_concurrency": settings.upload_concurrency, "verify_concurrency": settings.image_verify_concurrency,
"pending_grace_s": settings.pending_grace_seconds,
"fault_points": sorted(settings.fault_points)}})
@@ -543,6 +566,7 @@ def create_app(settings: Settings | None = None) -> Starlette:
for task in tasks:
task.cancel()
await sessions.close()
await staging.close()
await storage.close()
await db.close()
@@ -578,6 +602,7 @@ def create_app(settings: Settings | None = None) -> Starlette:
app.state.db = db
app.state.sessions = sessions
app.state.storage = storage
app.state.staging = staging
app.state.upload_slots = asyncio.Semaphore(settings.upload_concurrency)
app.state.verify_slots = asyncio.Semaphore(settings.image_verify_concurrency)
app.state.draining = False
+86 -47
View File
@@ -1,26 +1,29 @@
"""Attachments (lab 4): metadata in PostgreSQL, original bytes in the private S3 bucket.
"""Attachments: metadata in PostgreSQL, bytes in private S3 buckets (lab 4), processed
asynchronously by the worker (lab 5, docs/PIPELINE.md).
Upload (POST /chats/{id}/attachments, raw PNG/JPEG body), in this order:
auth -> membership (404) -> filename / Content-Type / Content-Length (400/415/413)
-> upload slot (at most UPLOAD_CONCURRENCY bodies per replica, else 503 + Retry-After)
-> body in chunks: size limit, read deadline, format + pixel count from the first bytes
-> full integrity check (images.verify) -> SHA-256 / size known
-> INSERT row status='pending' with a fresh random object_key, COMMIT (intent)
-> PUT object to S3 (Content-MD5)
-> UPDATE status='ready' WHERE status='pending', COMMIT -> 201 ready.
-> full integrity check (images.verify) -> SHA-256 / size known (415 is decided HERE)
-> INSERT row status='pending' with a fresh random staging key, COMMIT (intent)
-> PUT the original into the STAGING bucket (Content-MD5)
-> ONE transaction: status 'pending' -> 'queued' + jobs row + outbox row, COMMIT
-> 202 queued.
Why 201 is honest: it is sent only after BOTH the object and the ready row are durable.
What a failure between the steps leaves behind, and who removes it (docs/CONSISTENCY.md):
What is durable at 202: the complete, verified original in the staging bucket (S3 confirmed the
PUT and its MD5), the metadata row, the job and the outbox row that makes the relay publish it.
A crash of the API after the answer loses nothing; a broker that is down only delays the publish.
* before the intent COMMIT: nothing (no object was written);
* after the intent, before/while PUT, before the ready COMMIT: a 'pending' row (+ maybe the
object). The client got 503 and may retry (a new attachment). The janitor (below) deletes the
object - idempotent, a missing key is fine - and then the row, once it is older than
PENDING_GRACE_SECONDS. Nothing is compensated in the request path: after a lost answer of
COMMIT we do not know whether the row is ready, and deleting the object could destroy a
ready attachment. Only the janitor decides, under a row lock, from the committed state;
* after the ready COMMIT (answer lost): a correct ready attachment the client does not know.
Failures between the steps (docs/PIPELINE.md, failure windows):
* before the intent COMMIT: nothing was written;
* after the intent, before/while PUT, or between PUT and the job COMMIT (FAULT_POINTS=after_s3_put):
a 'pending' row (+ maybe the staging object), invisible to everybody (404). The client got 503
and may retry (a new attachment). The janitor deletes the staging object - idempotent - and then
the row once it is older than PENDING_GRACE_SECONDS, deciding under the row lock;
* after the job COMMIT, answer lost (FAULT_POINTS=after_enqueue_commit): a correct queued
attachment the client does not know; the worker processes it as usual.
"""
from __future__ import annotations
@@ -36,11 +39,11 @@ import psycopg
from psycopg import errors as pg_errors
from starlette.requests import Request
from . import images
from . import images, jobs
from .config import Settings
from .db import Database
from .errors import ApiError, InvalidRequest, NotFound, PayloadTooLarge, Unavailable, UnsupportedMediaType
from .storage import OBJECT_PREFIX, Storage
from .storage import STAGING_PREFIX, Storage
from .timefmt import rfc3339
log = logging.getLogger("messenger.attachments")
@@ -48,6 +51,21 @@ log = logging.getLogger("messenger.attachments")
JANITOR_LOCK_ID = 2026_04 # advisory lock: one janitor round at a time across all replicas
@dataclass(frozen=True, slots=True)
class Variant:
name: str
content_type: str
size_bytes: int
width: int
height: int
sha256: str
object_key: str
def to_json(self) -> dict:
return {"name": self.name, "content_type": self.content_type, "size_bytes": self.size_bytes,
"width": self.width, "height": self.height, "sha256": self.sha256}
@dataclass(frozen=True, slots=True)
class Attachment:
id: str
@@ -57,100 +75,121 @@ class Attachment:
content_type: str
size_bytes: int
sha256: str
object_key: str
object_key: str | None # published original in the final bucket (None until the worker publishes it)
status: str
created_at: dt.datetime
updated_at: dt.datetime
lost: bool = False
error_code: str | None = None
published_version: int | None = None
variants: tuple[Variant, ...] = ()
def to_json(self) -> dict:
"""Public representation: no object key, no bucket, no internal state."""
"""Public representation: no object keys, no buckets, no job internals."""
return {"id": self.id, "chat_id": self.chat_id, "owner_id": self.owner_id, "filename": self.filename,
"content_type": self.content_type, "size_bytes": self.size_bytes, "sha256": self.sha256,
"status": "ready", "variants": [], "error_code": None,
"status": self.status, "variants": [v.to_json() for v in self.variants] if self.status == "ready" else [],
"error_code": self.error_code if self.status == "failed" else None,
"created_at": rfc3339(self.created_at), "updated_at": rfc3339(self.updated_at)}
def variant(self, name: str) -> Variant | None:
return next((v for v in self.variants if v.name == name), None)
COLUMNS = ("a.id, a.chat_id, a.owner_id, a.filename, a.content_type, a.size_bytes, a.sha256, a.object_key, "
"a.status, a.created_at, a.updated_at, a.lost_at IS NOT NULL")
"a.status, a.created_at, a.updated_at, a.lost_at IS NOT NULL, a.error_code, a.published_version")
def _attachment(row) -> Attachment:
def _attachment(row, variants: tuple[Variant, ...] = ()) -> Attachment:
return Attachment(str(row[0]), str(row[1]), str(row[2]), row[3], row[4], row[5], bytes(row[6]).hex(), row[7],
row[8], row[9], row[10], row[11])
row[8], row[9], row[10], row[11], row[12], row[13], variants)
def new_object_key() -> str:
def new_staging_key() -> str:
"""Server-generated, 128 random bits, unrelated to the attachment id and the filename."""
return OBJECT_PREFIX + secrets.token_hex(16)
return STAGING_PREFIX + secrets.token_hex(16)
# ------------------------------------------------------------------ SQL
async def create_intent(db: Database, attachment_id: uuid.UUID, chat_id: str, owner_id: str, filename: str,
info: images.ImageInfo, size: int, sha256: bytes, object_key: str) -> None:
info: images.ImageInfo, size: int, sha256: bytes, staging_key: str) -> None:
async def work(conn: psycopg.AsyncConnection) -> None:
await conn.execute(
"INSERT INTO attachments (id, chat_id, owner_id, filename, content_type, size_bytes, sha256, "
"width, height, object_key, status) VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, 'pending')",
"width, height, staging_key, staging_state, status) "
"VALUES (%s, %s, %s, %s, %s, %s, %s, %s, %s, %s, 'present', 'pending')",
(attachment_id, chat_id, owner_id, filename, info.content_type, size, sha256, info.width, info.height,
object_key))
staging_key))
try:
await db.run(work, transaction=False)
except pg_errors.ForeignKeyViolation:
raise NotFound("Чат не найден") from None
async def mark_ready(db: Database, attachment_id: uuid.UUID) -> Attachment | None:
"""pending -> ready. None if the row is no longer pending (the janitor removed a stale intent)."""
async def enqueue(db: Database, attachment_id: uuid.UUID, settings: Settings) -> Attachment | None:
"""pending -> queued + job + outbox row in ONE transaction (the durable handoff before 202).
None if the row is no longer pending (the janitor removed a stale intent)."""
async def work(conn: psycopg.AsyncConnection) -> Attachment | None:
cur = await conn.execute(
"UPDATE attachments a SET status = 'ready', updated_at = clock_timestamp() "
"UPDATE attachments a SET status = 'queued', updated_at = clock_timestamp() "
f"WHERE a.id = %s AND a.status = 'pending' RETURNING {COLUMNS}", (attachment_id,))
row = await cur.fetchone()
return _attachment(row) if row else None
return await db.run(work, transaction=False)
if row is None:
return None
await jobs.create_job(conn, attachment_id, settings.pipeline_version, "upload", settings.job_max_attempts,
"api:" + settings.instance_id)
return _attachment(row)
return await db.run(work)
async def get_visible(db: Database, user_id: str, attachment_id: str) -> Attachment:
"""A READY attachment of a chat the user is a member of; anything else is 404 (unknown,
foreign chat, unfinished upload) - existence of foreign objects is never revealed."""
"""An attachment of a chat the user is a member of, in any state but the invisible intent
'pending'; anything else is 404 (unknown, foreign chat, unfinished upload) - existence of
foreign objects is never revealed. A ready attachment comes with its published variants."""
async def work(conn: psycopg.AsyncConnection) -> Attachment:
cur = await conn.execute(
f"SELECT {COLUMNS} FROM attachments a "
"JOIN chat_members m ON m.chat_id = a.chat_id AND m.user_id = %s "
"WHERE a.id = %s AND a.status = 'ready'", (user_id, attachment_id))
"WHERE a.id = %s AND a.status <> 'pending'", (user_id, attachment_id))
row = await cur.fetchone()
if row is None:
raise NotFound("Вложение не найдено")
return _attachment(row)
return await db.run(work, transaction=False)
variants: tuple[Variant, ...] = ()
if row[8] == "ready":
cur = await conn.execute(
"SELECT name, content_type, size_bytes, width, height, sha256, object_key FROM attachment_variants "
"WHERE attachment_id = %s AND pipeline_version = %s "
"ORDER BY array_position(ARRAY['thumbnail', 'crop', 'optimized'], name)", (row[0], row[13]))
variants = tuple(Variant(r[0], r[1], r[2], r[3], r[4], bytes(r[5]).hex(), r[6]) for r in await cur.fetchall())
return _attachment(row, variants)
# REPEATABLE READ is not needed: the variant rows of a published version never change.
return await db.run(work)
async def clean_stale_pending(db: Database, storage: Storage, grace_seconds: int, batch: int = 20) -> int:
"""Janitor round: remove intents of uploads that never became ready.
async def clean_stale_pending(db: Database, staging: Storage, grace_seconds: int, batch: int = 20) -> int:
"""Janitor round: remove intents of uploads that never got their job (no 202 was sent).
Under FOR UPDATE: an upload that is (impossibly) still running would block on the row in
mark_ready and then find it gone (-> 503), so a ready attachment can never lose its object.
Order: object first (idempotent), row second; a crash in between leaves the pending row and
the next round repeats. Only status='pending' rows older than the grace period are touched:
fresh uploads and ready/used attachments are never selected.
enqueue() and then find it gone (-> 503). Order: staging object first (idempotent), row second;
a crash in between leaves the pending row and the next round repeats. Only status='pending'
rows older than the grace period are touched: queued/processing/ready/failed never.
"""
async def work(conn: psycopg.AsyncConnection) -> int:
cur = await conn.execute("SELECT pg_try_advisory_xact_lock(%s)", (JANITOR_LOCK_ID,))
if not (await cur.fetchone())[0]:
return 0
cur = await conn.execute(
"SELECT id, object_key FROM attachments WHERE status = 'pending' "
"SELECT id, staging_key FROM attachments WHERE status = 'pending' "
"AND created_at < clock_timestamp() - make_interval(secs => %s) "
"ORDER BY created_at LIMIT %s FOR UPDATE SKIP LOCKED", (grace_seconds, batch))
rows = await cur.fetchall()
if not rows:
return 0
await asyncio.gather(*(storage.delete(key) for _, key in rows))
await asyncio.gather(*(staging.delete(key) for _, key in rows))
await conn.execute("DELETE FROM attachments WHERE id = ANY(%s) AND status = 'pending'",
([r[0] for r in rows],))
return len(rows)
return await db.run(work, deadline=storage.timeout + 5)
return await db.run(work, deadline=staging.timeout + 5)
# ------------------------------------------------------------------ request body
+3 -3
View File
@@ -55,7 +55,7 @@ def _forget(task: asyncio.Task) -> None:
class Database:
def __init__(self, settings: Settings):
def __init__(self, settings: Settings, name: str = "messenger-api"):
self.deadline = settings.db_timeout_seconds
options = (
f"-c statement_timeout={settings.db_statement_timeout_ms} "
@@ -68,12 +68,12 @@ class Database:
max_size=settings.db_pool_max,
timeout=settings.db_timeout_seconds, # max wait for a free connection
open=False,
name="messenger",
name=name,
check=AsyncConnectionPool.check_connection, # drop connections broken by a DB restart
kwargs={
"options": options,
"connect_timeout": max(1, int(settings.db_timeout_seconds)),
"application_name": "messenger-api",
"application_name": name, # visible in pg_stat_activity: api / relay / worker
# Autocommit connections: a single statement (post_message function, ping) is its
# own transaction with ONE round trip; multi-statement work uses an explicit
# conn.transaction() (BEGIN ... COMMIT), see run().
+8 -7
View File
@@ -230,11 +230,12 @@ async def _insert_message(conn: psycopg.AsyncConnection, message_id: uuid.UUID,
return created_at
# Attachments a message may reference: ready, of THIS chat (else 404), not quarantined (else 409).
# FOR SHARE keeps a concurrent quarantine (reconciler) from slipping in before COMMIT.
# Attachments a message may reference (lab 5): of THIS chat and not the invisible 'pending' intent
# (else 404); queued/processing/ready are usable, failed or quarantined -> 409.
# FOR SHARE keeps a concurrent transition to failed from slipping in before COMMIT.
CHECK_ATTACHMENTS = (
"SELECT count(*), count(*) FILTER (WHERE a.lost_at IS NULL) FROM ("
"SELECT lost_at FROM attachments WHERE id = ANY(%s::uuid[]) AND chat_id = %s AND status = 'ready' "
"SELECT count(*), count(*) FILTER (WHERE a.lost_at IS NULL AND a.status <> 'failed') FROM ("
"SELECT lost_at, status FROM attachments WHERE id = ANY(%s::uuid[]) AND chat_id = %s AND status <> 'pending' "
"FOR SHARE) a")
@@ -280,7 +281,7 @@ async def post_message(db: Database, user_id: str, chat_id: str, key: str, reque
ever left behind: a crash or 5xx rolls everything back and the client may retry);
* same key, same body (any replica, also concurrently) -> the stored Message, 201, replayed;
* same key, other body -> 409; the membership check comes first, so outsiders always get 404;
* attachments (lab 4) are checked before the key is claimed (same order as post_message_v2).
* attachments (lab 4) are checked before the key is claimed (same order as post_message_v3).
"""
message_id = uuid.uuid4()
attachment_ids = list(attachment_ids)
@@ -308,12 +309,12 @@ async def post_message(db: Database, user_id: str, chat_id: str, key: str, reque
POST_MESSAGE_CALL = ("SELECT o_status, o_id, o_text, o_created_at, o_attachments "
"FROM post_message_v2(%s, %s, %s, %s, %s, %s, %s, %s::uuid[])")
"FROM post_message_v3(%s, %s, %s, %s, %s, %s, %s, %s::uuid[])")
async def post_message_single_call(db: Database, user_id: str, chat_id: str, key: str, request_hash: bytes,
text: str, ttl_seconds: int, attachment_ids: list[str] = ()) -> PostResult:
"""Same contract as post_message, ONE round trip: SQL function post_message_v2 (migration 0004 =
"""Same contract as post_message, ONE round trip: SQL function post_message_v3 (migration 0005 =
0003 + attachment_ids) in autocommit mode, i.e. the call is the whole transaction."""
message_id = uuid.uuid4()
+14 -3
View File
@@ -58,8 +58,10 @@ class ListedObject:
class Storage:
def __init__(self, settings: Settings, *, access_key: str | None = None, secret_key: str | None = None):
self.bucket = settings.s3_bucket
def __init__(self, settings: Settings, *, access_key: str | None = None, secret_key: str | None = None,
bucket: str | None = None):
# Lab 5: one client per bucket - the final bucket (S3_BUCKET) or the staging bucket.
self.bucket = bucket or settings.s3_bucket
self.timeout = settings.s3_timeout_seconds
self._endpoint = settings.s3_endpoint
self._region = settings.s3_region
@@ -174,4 +176,13 @@ class Storage:
SENTINEL_KEY = "health/sentinel"
OBJECT_PREFIX = "o/" # every attachment original lives under o/<random 128-bit hex>
# Lab 4 originals: o/<random 128-bit hex> in the final bucket (still readable).
OBJECT_PREFIX = "o/"
# Lab 5: s/<random 128-bit hex> in the STAGING bucket (written by the API before 202) and
# a/<attachment id>/<attempt token>/<original|thumbnail|crop|optimized> in the final bucket
# (written by the worker; the token makes every attempt write its own keys).
STAGING_PREFIX = "s/"
def attempt_key(attachment_id: str, token: str, name: str) -> str:
return f"a/{attachment_id}/{token}/{name}"
+6 -10
View File
@@ -181,19 +181,15 @@ def filename_param(query: list[tuple[str, str]]) -> str:
return value
ATTACHMENT_VARIANTS = frozenset({"thumbnail", "crop", "optimized"}) # exist in the schema, not in lab 4
ATTACHMENT_VARIANTS = ("original", "thumbnail", "crop", "optimized")
def variant_param(query: list[tuple[str, str]]) -> str:
"""`variant` of GET /attachments/{id}/content: only `original` exists in lab 4. A known variant
name is a missing variant of a ready object (404); anything else is an invalid value (400)."""
"""`variant` of GET /attachments/{id}/content: original (default) or a lab 5 variant name.
Any other value is 400; a known name that a ready attachment lacks is 404 (handler)."""
value = single_query_param(query, "variant")
if value is None or value == "original":
if value is None:
return "original"
if value in ATTACHMENT_VARIANTS:
raise NotFoundVariant()
raise InvalidRequest("variant должен быть original")
class NotFoundVariant(Exception):
"""Raised for a variant that lab 4 does not produce; mapped to 404 by the handler."""
return value
raise InvalidRequest("variant должен быть original, thumbnail, crop или optimized")