lab4: вложения в приватном S3 (SeaweedFS), метаданные в PostgreSQL

- compose: сервис s3 (SeaweedFS 4.46 по digest, volume s3data, внутренняя сеть storage,
  web-интерфейсы только на 127.0.0.1 внутри контейнера), одноразовый s3-init;
- три пары ключей S3: admin (только bootstrap), api (Read+Write своего bucket),
  reconciler (+List); bootstrap идемпотентен и проверяет приватность и права;
- миграция 0004: attachments (pending -> ready, случайный object_key), message_attachments,
  post_message_v2 с проверкой вложений в той же транзакции;
- POST /chats/{id}/attachments: лимиты до чтения/декодирования, потоковая проверка PNG
  (CRC + zlib до точного размера), JPEG через djpeg -strict, семафор загрузок;
- GET метаданных и /content (стриминг из S3, без redirect), readiness с S3,
  HAProxy: деградированный режим без S3 через /health/ready/core;
- janitor незавершённых загрузок с grace period.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Alexey Bodanyuk
2026-10-25 13:26:01 +03:00
co-authored by Claude Opus 5.5
parent d916d09f7b
commit 335df4b2b2
19 changed files with 2586 additions and 69 deletions
+160 -12
View File
@@ -6,6 +6,7 @@ import contextlib
import hashlib
import json
import logging
import random
import re
import time
import uuid
@@ -14,15 +15,16 @@ from typing import Any
from starlette.applications import Starlette
from starlette.exceptions import HTTPException
from starlette.requests import Request
from starlette.responses import JSONResponse, Response
from starlette.responses import JSONResponse, Response, StreamingResponse
from starlette.routing import Route
from starlette.types import ASGIApp, Message as AsgiMessage, Receive, Scope, Send
from . import authhdr, cursor, docs, passwords, repo, validation
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 .sessions import SessionStore, build_store
from .storage import ObjectMissing, Storage
from .timefmt import rfc3339
log = logging.getLogger("messenger")
@@ -126,6 +128,10 @@ def sessions_of(request: Request) -> SessionStore:
return request.app.state.sessions
def storage_of(request: Request) -> Storage:
return request.app.state.storage
async def authenticate(request: Request) -> str:
"""User id of a valid Bearer session, otherwise 401. X-User-Id is ignored since lab 2.
@@ -206,15 +212,32 @@ async def live(request: Request) -> JSONResponse:
async def ready(request: Request) -> JSONResponse:
"""200 only if the mandatory synchronous path can be served: PostgreSQL and,
with SESSION_STORE=valkey, Valkey answer. 503 while draining after SIGTERM."""
"""200 only if the WHOLE mandatory synchronous path can be served: PostgreSQL, Valkey
(SESSION_STORE=valkey) and, since lab 4, the S3 bucket (upload/download of attachments)
answer. 503 while draining after SIGTERM. HAProxy routes to such replicas first."""
if request.app.state.draining:
raise Unavailable("Сервер завершает работу")
db_ok, store_ok = await asyncio.gather(db_of(request).ping(), sessions_of(request).ping())
db_ok, store_ok, s3_ok = await asyncio.gather(db_of(request).ping(), sessions_of(request).ping(),
storage_of(request).ping())
if not db_ok:
raise Unavailable("PostgreSQL недоступен")
if not store_ok:
raise Unavailable("Хранилище сессий недоступно")
if not s3_ok:
raise Unavailable("Хранилище файлов (S3) недоступно")
return ok({"status": "ok"})
async def ready_core(request: Request) -> JSONResponse:
"""Not in the contract: readiness WITHOUT S3, for the load balancer's degraded mode only.
If S3 is down for every replica, /health/ready is 503 everywhere; HAProxy then routes to
replicas that are at least core-ready, so chats and messages keep working while every
attachment request answers 503 itself (see infra/haproxy/haproxy.cfg)."""
if request.app.state.draining:
raise Unavailable("Сервер завершает работу")
db_ok, store_ok = await asyncio.gather(db_of(request).ping(), sessions_of(request).ping())
if not (db_ok and store_ok):
raise Unavailable("PostgreSQL или хранилище сессий недоступно")
return ok({"status": "ok"})
@@ -288,12 +311,12 @@ async def get_chat(request: Request) -> JSONResponse:
def parse_message(data: dict[str, Any]) -> tuple[str, list[str]]:
"""`text` is required and a string (also with attachments); `attachment_ids`, if present, an
array of 0-10 unique UUIDs. "" or whitespace-only text is allowed only with attachments."""
validation.check_fields(data, {"text"}, frozenset({"attachment_ids"}))
text = validation.text_field(data, "text", 4000)
attachment_ids = validation.attachment_ids_field(data)
if attachment_ids:
# Lab 3 has no attachments: every referenced ID is unknown (lab 4 semantics: 404).
raise NotFound("Вложение не найдено")
text = validation.text_field(data, "text", 4000, min_len=0 if attachment_ids else 1,
allow_blank=bool(attachment_ids))
return text, attachment_ids
@@ -309,7 +332,7 @@ async def create_message(request: Request) -> JSONResponse:
key = idempotency_key(request)
text, attachment_ids = parse_message(await read_json(request))
result = await repo.post_message(db, user_id, chat_id, key, request_fingerprint(text, attachment_ids), text,
settings.idempotency_ttl_seconds)
settings.idempotency_ttl_seconds, attachment_ids)
else:
try:
key = idempotency_key(request)
@@ -321,7 +344,7 @@ async def create_message(request: Request) -> JSONResponse:
# Membership check, key claim, chat clock and message: one SQL call, one transaction.
result = await repo.post_message_single_call(db, user_id, chat_id, key,
request_fingerprint(text, attachment_ids), text,
settings.idempotency_ttl_seconds)
settings.idempotency_ttl_seconds, attachment_ids)
if result.replayed:
log.info("idempotent_replay", extra={"fields": {"request_id": request.state.request_id,
"message_id": result.message.id}})
@@ -340,10 +363,111 @@ async def list_messages(request: Request) -> JSONResponse:
return ok({"items": [m.to_json() for m in items], "next_cursor": next_cursor})
# ------------------------------------------------------------------ attachments (lab 4)
def fault(request: Request, point: str) -> None:
"""Failure injection for the level 4 experiments (FAULT_POINTS, compose.faults.yaml only)."""
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)."""
settings: Settings = request.app.state.settings
reader = attachments.BodyReader(request)
try:
user_id = await authenticate(request)
chat_id = path_uuid(request, "chat_id")
db = db_of(request)
await repo.ensure_member(db, user_id, chat_id) # outsiders: 404 before anything about the body
filename = validation.filename_param(validation.query_items(request.scope["query_string"]))
declared = attachments.declared_type(request)
attachments.declared_length(request, settings.max_upload_bytes)
slots: asyncio.Semaphore = request.app.state.upload_slots
try:
await asyncio.wait_for(slots.acquire(), settings.upload_queue_timeout_seconds)
except asyncio.TimeoutError:
raise attachments.Busy("Слишком много одновременных загрузок, повторите позже") from None
try:
upload = await attachments.read_image(reader, settings, declared)
finally:
slots.release()
except ApiError:
# Early rejection: drop (not store) a bounded rest of the body so the client reads our error.
await reader.discard(settings.max_upload_bytes + (1 << 20), timeout=5.0)
raise
attachment_id = uuid.uuid4()
object_key = attachments.new_object_key()
storage = storage_of(request)
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()})
fault(request, "after_s3_put")
attachment = await attachments.mark_ready(db, attachment_id)
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)
async def get_attachment(request: Request) -> JSONResponse:
user_id = await authenticate(request)
attachment_id = path_uuid(request, "attachment_id")
return ok((await attachments.get_visible(db_of(request), user_id, attachment_id)).to_json())
def content_disposition(filename: str) -> str:
"""RFC 6266/5987: ASCII fallback + UTF-8 name; the stored name itself is never interpreted."""
from urllib.parse import quote
fallback = "".join(c if 32 <= ord(c) < 127 and c not in '"\\%;' else "_" for c in filename)
return f"attachment; filename=\"{fallback}\"; filename*=UTF-8''{quote(filename, safe='')}"
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)."""
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:
await attachments.get_visible(db_of(request), user_id, attachment_id) # 404 for outsiders first
raise NotFound("Такого варианта у вложения нет") from None
attachment = await attachments.get_visible(db_of(request), user_id, attachment_id)
if attachment.lost:
raise NotFound("Оригинал вложения утрачен")
try:
size, chunks = await storage_of(request).open_stream(attachment.object_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:
log.error("attachment_size_mismatch", extra={"fields": {"request_id": request.state.request_id,
"attachment_id": attachment.id}})
await chunks.aclose()
raise Unavailable("Оригинал вложения временно недоступен")
headers = {
"Content-Length": str(size),
"Content-Disposition": content_disposition(attachment.filename),
"X-Content-Type-Options": "nosniff",
"Cache-Control": "private, no-store",
"ETag": '"' + attachment.sha256 + '"',
}
return StreamingResponse(chunks, media_type=attachment.content_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)
async def purge_loop() -> None:
"""Housekeeping: expired Idempotency-Key rows (older than IDEMPOTENCY_TTL_SECONDS)."""
@@ -356,6 +480,18 @@ def create_app(settings: Settings | None = None) -> Starlette:
except ApiError:
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).
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)
if removed:
log.info("stale_uploads_removed", extra={"fields": {"count": removed}})
except ApiError:
pass # PostgreSQL or S3 unavailable right now; the rows stay pending for the next round
stats_log = logging.getLogger("messenger.stats")
async def stats_loop(every: float) -> None:
@@ -376,12 +512,17 @@ def create_app(settings: Settings | None = None) -> Starlette:
async def lifespan(app: Starlette):
passwords.configure(settings.hash_concurrency)
await db.open()
await storage.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}})
"message_write_path": settings.message_write_path, "s3_bucket": settings.s3_bucket,
"upload_concurrency": settings.upload_concurrency, "pending_grace_s": settings.pending_grace_seconds,
"fault_points": sorted(settings.fault_points)}})
tasks = [asyncio.create_task(purge_loop())]
if settings.janitor_interval_seconds > 0:
tasks.append(asyncio.create_task(janitor_loop(settings.janitor_interval_seconds)))
if settings.stats_log_seconds > 0:
tasks.append(asyncio.create_task(stats_loop(settings.stats_log_seconds)))
try:
@@ -390,11 +531,13 @@ def create_app(settings: Settings | None = None) -> Starlette:
for task in tasks:
task.cancel()
await sessions.close()
await storage.close()
await db.close()
routes = [
Route("/health/live", live, methods=["GET"]),
Route("/health/ready", ready, methods=["GET"]),
Route("/health/ready/core", ready_core, methods=["GET"]),
Route("/api/v1/users", register_user, methods=["POST"]),
Route("/api/v1/users/me", me, methods=["GET"]),
Route("/api/v1/auth/sessions", create_session, methods=["POST"]),
@@ -404,6 +547,9 @@ def create_app(settings: Settings | None = None) -> Starlette:
Route("/api/v1/chats/{chat_id}", get_chat, methods=["GET"]),
Route("/api/v1/chats/{chat_id}/messages", create_message, methods=["POST"]),
Route("/api/v1/chats/{chat_id}/messages", list_messages, methods=["GET"]),
Route("/api/v1/chats/{chat_id}/attachments", upload_attachment, methods=["POST"]),
Route("/api/v1/attachments/{attachment_id}", get_attachment, methods=["GET"]),
Route("/api/v1/attachments/{attachment_id}/content", download_attachment, methods=["GET"]),
]
if settings.docs_enabled: # Swagger UI for manual experiments, not part of the contract
routes += [
@@ -419,5 +565,7 @@ def create_app(settings: Settings | None = None) -> Starlette:
app.state.settings = settings
app.state.db = db
app.state.sessions = sessions
app.state.storage = storage
app.state.upload_slots = asyncio.Semaphore(settings.upload_concurrency)
app.state.draining = False
return app
+285
View File
@@ -0,0 +1,285 @@
"""Attachments (lab 4): metadata in PostgreSQL, original bytes in the private S3 bucket.
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.
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):
* 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.
"""
from __future__ import annotations
import asyncio
import datetime as dt
import hashlib
import logging
import secrets
import uuid
from dataclasses import dataclass
import psycopg
from psycopg import errors as pg_errors
from starlette.requests import Request
from . import images
from .config import Settings
from .db import Database
from .errors import ApiError, InvalidRequest, NotFound, PayloadTooLarge, Unavailable, UnsupportedMediaType
from .storage import OBJECT_PREFIX, Storage
from .timefmt import rfc3339
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 Attachment:
id: str
chat_id: str
owner_id: str
filename: str
content_type: str
size_bytes: int
sha256: str
object_key: str
status: str
created_at: dt.datetime
updated_at: dt.datetime
lost: bool = False
def to_json(self) -> dict:
"""Public representation: no object key, no bucket, no internal state."""
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,
"created_at": rfc3339(self.created_at), "updated_at": rfc3339(self.updated_at)}
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")
def _attachment(row) -> 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])
def new_object_key() -> str:
"""Server-generated, 128 random bits, unrelated to the attachment id and the filename."""
return OBJECT_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:
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')",
(attachment_id, chat_id, owner_id, filename, info.content_type, size, sha256, info.width, info.height,
object_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 work(conn: psycopg.AsyncConnection) -> Attachment | None:
cur = await conn.execute(
"UPDATE attachments a SET status = 'ready', 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)
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."""
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))
row = await cur.fetchone()
if row is None:
raise NotFound("Вложение не найдено")
return _attachment(row)
return await db.run(work, transaction=False)
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.
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.
"""
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' "
"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 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)
# ------------------------------------------------------------------ request body
class Busy(Unavailable):
"""All upload slots of this replica are taken: retry shortly (no body was read)."""
@property
def headers(self) -> dict[str, str]:
return {"Retry-After": "1"}
@dataclass
class Upload:
data: bytes
info: images.ImageInfo
sha256: bytes
class BodyReader:
"""Reads the raw ASGI body ourselves (not request.stream()), so we always know how much of it
is consumed and can discard the rest after an early rejection."""
def __init__(self, request: Request):
self.receive = request.receive
self.done = False
async def chunk(self, timeout: float) -> bytes | None:
if self.done:
return None
message = await asyncio.wait_for(self.receive(), timeout)
if message["type"] == "http.disconnect":
self.done = True
raise InvalidRequest("Клиент прервал загрузку")
if not message.get("more_body", False):
self.done = True
return message.get("body", b"")
async def discard(self, limit: int, timeout: float) -> None:
"""After an early 4xx: read and DROP up to `limit` more bytes, so an ordinary client that
writes the whole body before reading the answer still gets our JSON error instead of a
connection reset. Nothing is kept in memory; beyond the limit the connection is dropped."""
loop = asyncio.get_running_loop()
deadline = loop.time() + timeout
seen = 0
try:
while not self.done and seen <= limit:
remaining = deadline - loop.time()
if remaining <= 0:
return
seen += len(await self.chunk(remaining) or b"")
except (ApiError, asyncio.TimeoutError):
return
def declared_type(request: Request) -> str:
"""Content-Type of the upload must be image/png or image/jpeg (parameters ignored)."""
media_type = request.headers.get("content-type", "").partition(";")[0].strip().lower()
if media_type not in (images.PNG, images.JPEG):
raise UnsupportedMediaType("Content-Type должен быть image/png или image/jpeg")
return media_type
def declared_length(request: Request, limit: int) -> int | None:
raw = request.headers.get("content-length")
if raw is None:
return None # chunked transfer: the limit is enforced while reading
if not raw.isdigit():
raise InvalidRequest("Неверный Content-Length")
length = int(raw)
if length > limit:
raise PayloadTooLarge(f"Файл больше {limit} байт")
if length == 0:
raise InvalidRequest("Тело запроса пустое: нужны байты изображения")
return length
async def read_image(reader: BodyReader, settings: Settings, declared: str) -> Upload:
"""Body -> verified image. Memory: at most MAX_UPLOAD_BYTES per upload (bytearray); the format
and the pixel count are known after the first bytes, so a picture that is too big or not a
PNG/JPEG is rejected without reading/decoding the rest."""
loop = asyncio.get_running_loop()
deadline = loop.time() + settings.upload_read_timeout_seconds
limit = settings.max_upload_bytes
buffer = bytearray()
digest = hashlib.sha256()
info: images.ImageInfo | None = None
while True:
remaining = deadline - loop.time()
if remaining <= 0:
raise InvalidRequest("Тело запроса не получено вовремя")
try:
chunk = await reader.chunk(remaining)
except asyncio.TimeoutError:
raise InvalidRequest("Тело запроса не получено вовремя") from None
if chunk is None:
break
if len(buffer) + len(chunk) > limit:
raise PayloadTooLarge(f"Файл больше {limit} байт")
buffer += chunk
digest.update(chunk)
if info is None and buffer:
info = _sniff(buffer, settings, declared)
if not buffer:
raise InvalidRequest("Тело запроса пустое: нужны байты изображения")
if info is None:
raise UnsupportedMediaType("Файл не является корректным изображением PNG или JPEG")
data = bytes(buffer)
del buffer
try:
await images.verify(data, info, timeout=settings.image_verify_timeout_seconds,
jpeg_max_memory_mb=settings.jpeg_max_memory_mb)
except images.ImageError as exc:
raise _image_error(exc) from None
except asyncio.TimeoutError:
raise UnsupportedMediaType("Изображение не удалось проверить за отведённое время") from None
return Upload(data, info, digest.digest())
def _sniff(buffer: bytearray, settings: Settings, declared: str) -> images.ImageInfo | None:
try:
info = images.sniff(buffer, settings.max_image_pixels)
except images.ImageError as exc:
raise _image_error(exc) from None
if info is not None and info.content_type != declared:
# "MIME spoofing": the header says PNG, the bytes are JPEG (or vice versa).
raise UnsupportedMediaType("Содержимое не совпадает с Content-Type")
return info
def _image_error(exc: images.ImageError) -> ApiError:
return PayloadTooLarge(exc.message) if exc.status == 413 else UnsupportedMediaType(exc.message)
+64
View File
@@ -115,6 +115,25 @@ class Settings:
idempotency_ttl_seconds: int = 86400
stats_log_seconds: float = 0.0
message_write_path: str = "function"
# Lab 4: private S3 bucket and attachment limits (see storage.py, attachments.py).
s3_endpoint: str = "http://s3:8333"
s3_region: str = "us-east-1"
s3_bucket: str = "messenger-attachments"
s3_access_key_id: str = ""
s3_secret_access_key: str = ""
s3_timeout_seconds: float = 3.0
s3_connect_timeout_seconds: float = 1.0
s3_max_connections: int = 20
max_upload_bytes: int = 10 * 1024 * 1024
max_image_pixels: int = 40_000_000
upload_concurrency: int = 4
upload_queue_timeout_seconds: float = 5.0
upload_read_timeout_seconds: float = 30.0
image_verify_timeout_seconds: float = 10.0
jpeg_max_memory_mb: int = 256
pending_grace_seconds: int = 900
janitor_interval_seconds: float = 60.0
fault_points: frozenset[str] = frozenset()
@property
def db_statement_timeout_ms(self) -> int:
@@ -165,4 +184,49 @@ class Settings:
# Track 1: `function` = one SQL call per send (migration 0003), `multi` = the previous
# multi-statement transaction; both have identical semantics (A/B on one image).
message_write_path=_choice("MESSAGE_WRITE_PATH", "function", {"function", "multi"}),
**s3_settings(),
)
# Contract limits (contracts/README.md): configurable DOWN for experiments, never above them.
CONTRACT_MAX_UPLOAD_BYTES = 10 * 1024 * 1024
CONTRACT_MAX_IMAGE_PIXELS = 40_000_000
# Failure points for the level 4 experiments only (compose.faults.yaml); empty in normal runs.
FAULT_POINTS = {"after_s3_put", "after_ready_commit"}
def s3_settings(*, credentials_required: bool = True) -> dict:
"""S3 part of Settings; also used by the reconciler/bootstrap CLIs with their own keys."""
faults = frozenset(f.strip() for f in os.environ.get("FAULT_POINTS", "").split(",") if f.strip())
if faults - FAULT_POINTS:
raise ConfigError(f"FAULT_POINTS may contain only {sorted(FAULT_POINTS)}")
endpoint = os.environ.get("S3_ENDPOINT", "http://s3:8333").strip() or "http://s3:8333"
if not re.fullmatch(r"https?://[A-Za-z0-9.\-]+(:[0-9]{1,5})?/?", endpoint):
raise ConfigError(f"S3_ENDPOINT must be http(s)://host[:port], got {endpoint!r}")
bucket = os.environ.get("S3_BUCKET", "messenger-attachments").strip() or "messenger-attachments"
if not re.fullmatch(r"[a-z0-9][a-z0-9.-]{1,61}[a-z0-9]", bucket):
raise ConfigError(f"S3_BUCKET is not a valid bucket name: {bucket!r}")
return dict(
s3_endpoint=endpoint,
s3_region=os.environ.get("S3_REGION", "us-east-1").strip() or "us-east-1",
s3_bucket=bucket,
s3_access_key_id=_secret("S3_ACCESS_KEY_ID", required=credentials_required),
s3_secret_access_key=_secret("S3_SECRET_ACCESS_KEY", required=credentials_required),
s3_timeout_seconds=_float("S3_TIMEOUT_SECONDS", 3.0, 0.2, 30.0),
s3_connect_timeout_seconds=_float("S3_CONNECT_TIMEOUT_SECONDS", 1.0, 0.1, 10.0),
s3_max_connections=_int("S3_MAX_CONNECTIONS", 20, 1, 200),
max_upload_bytes=_int("MAX_UPLOAD_BYTES", CONTRACT_MAX_UPLOAD_BYTES, 1024, CONTRACT_MAX_UPLOAD_BYTES),
max_image_pixels=_int("MAX_IMAGE_PIXELS", CONTRACT_MAX_IMAGE_PIXELS, 1, CONTRACT_MAX_IMAGE_PIXELS),
# At most this many request bodies of one replica are buffered/checked at once (RAM bound:
# UPLOAD_CONCURRENCY x MAX_UPLOAD_BYTES); further uploads wait, then get 503.
upload_concurrency=_int("UPLOAD_CONCURRENCY", 4, 1, 64),
upload_queue_timeout_seconds=_float("UPLOAD_QUEUE_TIMEOUT_SECONDS", 5.0, 0.0, 60.0),
upload_read_timeout_seconds=_float("UPLOAD_READ_TIMEOUT_SECONDS", 30.0, 1.0, 300.0),
image_verify_timeout_seconds=_float("IMAGE_VERIFY_TIMEOUT_SECONDS", 10.0, 1.0, 120.0),
jpeg_max_memory_mb=_int("JPEG_MAX_MEMORY_MB", 256, 16, 2048),
# An intent row still `pending` after this long belongs to a failed upload: the janitor
# deletes its object and the row. Far above the longest possible upload (~45 s).
pending_grace_seconds=_int("PENDING_GRACE_SECONDS", 900, 1, 7 * 24 * 3600),
janitor_interval_seconds=_float("JANITOR_INTERVAL_SECONDS", 60.0, 0.0, 3600.0),
fault_points=faults,
)
+6
View File
@@ -59,6 +59,12 @@ class PayloadTooLarge(ApiError):
default_message = "Тело запроса слишком большое"
class UnsupportedMediaType(ApiError):
status = 415
code = "unsupported_media_type"
default_message = "Неподдерживаемый тип содержимого"
class Unavailable(ApiError):
status = 503
code = "unavailable"
+317
View File
@@ -0,0 +1,317 @@
"""Content checks of uploaded images (lab 4): what the bytes ARE, not what the client says.
Two stages, both with bounded memory and both BEFORE anything is stored:
1. `sniff(head)` — runs on the first bytes while the body is still arriving. Recognises
PNG/JPEG by the signature and reads the dimensions from the header (PNG IHDR, JPEG SOFn).
A foreign format is rejected (415) and a picture with more than MAX_IMAGE_PIXELS pixels
(413) before the rest of the body is read and long before any pixel is decoded.
2. `verify(data, info)` — full integrity check of the complete file, run in a worker thread:
* PNG: our own streaming validator. Every chunk CRC is checked, the IDAT zlib stream is
inflated in 1 MiB steps and DISCARDED (no raster in memory), its size must be exactly
height x (1 + row bytes) (Adam7 per pass), every row filter byte must be 0..4, IEND
must close the file. A truncated file, a broken IDAT or extra/missing pixel data fail.
Pillow is not used for PNG: it would allocate the whole raster (40 Mpx RGBA = 160 MB).
* JPEG: libjpeg-turbo `djpeg -strict` decodes the whole entropy-coded data at 1/8 scale
(DCT scaling: the output raster is 64 times smaller) in a child process; any decoder
warning is fatal, so a truncated file or a file without scan data after SOS fails.
Every problem of the CONTENT is an ImageError carrying the HTTP status (415 or 413);
nothing here may turn user input into a 500.
"""
from __future__ import annotations
import asyncio
import os
import shutil
import struct
import zlib
from dataclasses import dataclass
PNG_SIGNATURE = b"\x89PNG\r\n\x1a\n"
JPEG_SOI = b"\xff\xd8\xff"
PNG = "image/png"
JPEG = "image/jpeg"
# Bits per pixel channel count by PNG colour type, and the bit depths allowed for it.
PNG_CHANNELS = {0: 1, 2: 3, 3: 1, 4: 2, 6: 4}
PNG_DEPTHS = {0: {1, 2, 4, 8, 16}, 2: {8, 16}, 3: {1, 2, 4, 8}, 4: {8, 16}, 6: {8, 16}}
# Adam7 passes: (x start, y start, x step, y step).
ADAM7 = ((0, 0, 8, 8), (4, 0, 8, 8), (0, 4, 4, 8), (2, 0, 4, 4), (0, 2, 2, 4), (1, 0, 2, 2), (0, 1, 1, 2))
# JPEG start-of-frame markers that carry the frame size (not DHT C4, JPG C8, DAC CC).
JPEG_SOF = {0xC0, 0xC1, 0xC2, 0xC3, 0xC5, 0xC6, 0xC7, 0xC9, 0xCA, 0xCB, 0xCD, 0xCE, 0xCF}
# Standalone markers without a length field.
JPEG_STANDALONE = {0x01, *range(0xD0, 0xD8)}
DJPEG = shutil.which("djpeg") or "djpeg" # libjpeg-turbo-progs, installed in the Dockerfile
INFLATE_STEP = 1 << 20 # bytes of raw PNG data produced per zlib call
MAX_DIMENSION = 2**31 - 1
class ImageError(Exception):
"""The uploaded bytes are not an acceptable image. `status`: 415 (format/corrupt) or 413."""
def __init__(self, status: int, message: str):
super().__init__(message)
self.status = status
self.message = message
def unsupported(message: str = "Файл не является корректным изображением PNG или JPEG") -> ImageError:
return ImageError(415, message)
@dataclass(frozen=True, slots=True)
class ImageInfo:
content_type: str
width: int
height: int
def detect(head: bytes) -> str | None:
"""Format by signature; None while fewer bytes than a signature have arrived."""
if head[:8] == PNG_SIGNATURE:
return PNG
if head[:3] == JPEG_SOI:
return JPEG
probe = min(len(head), 8)
if PNG_SIGNATURE.startswith(head[:probe]) or JPEG_SOI.startswith(head[:min(probe, 3)]):
return None
raise unsupported()
def sniff(head: bytes, max_pixels: int) -> ImageInfo | None:
"""Format and size from the beginning of the file, or None if more bytes are needed.
Raises ImageError: 415 for a foreign/broken header, 413 for too many pixels."""
kind = detect(head)
if kind is None:
return None
info = _png_header(head) if kind == PNG else _jpeg_header(head)
if info is None:
return None
if info.width * info.height > max_pixels:
raise ImageError(413, f"Изображение больше {max_pixels} пикселей")
return info
def _png_header(head: bytes) -> ImageInfo | None:
if len(head) < 33:
return None
length, ctype = struct.unpack(">I4s", head[8:16])
if ctype != b"IHDR" or length != 13:
raise unsupported()
if zlib.crc32(head[12:29]) != struct.unpack(">I", head[29:33])[0]:
raise unsupported()
width, height, depth, color, compression, filtering, interlace = struct.unpack(">IIBBBBB", head[16:29])
if not (1 <= width <= MAX_DIMENSION and 1 <= height <= MAX_DIMENSION):
raise unsupported()
if color not in PNG_DEPTHS or depth not in PNG_DEPTHS[color] or compression or filtering or interlace > 1:
raise unsupported()
return ImageInfo(PNG, width, height)
def _jpeg_header(head: bytes) -> ImageInfo | None:
"""Walk the marker segments up to the first SOFn. None = the SOF has not arrived yet."""
i = 2
n = len(head)
while True:
if i >= n:
return None
if head[i] != 0xFF:
raise unsupported()
while i < n and head[i] == 0xFF: # fill bytes
i += 1
if i >= n:
return None
marker = head[i]
i += 1
if marker in JPEG_STANDALONE:
continue
if marker in (0xD8, 0xD9, 0xDA, 0x00):
# Second SOI, EOI or scan data before any frame header: not a usable JPEG.
raise unsupported()
if i + 2 > n:
return None
length = struct.unpack(">H", head[i:i + 2])[0]
if length < 2:
raise unsupported()
if marker in JPEG_SOF:
if i + 7 > n:
return None
if length < 8:
raise unsupported()
height, width = struct.unpack(">HH", head[i + 3:i + 7])
if width == 0 or height == 0: # height 0 = DNL-defined height: not supported
raise unsupported()
return ImageInfo(JPEG, width, height)
i += length
# ------------------------------------------------------------------ full verification
async def verify(data: bytes | bytearray | memoryview, info: ImageInfo, *, timeout: float = 10.0,
jpeg_max_memory_mb: int = 256) -> None:
"""Complete integrity check of the whole file. PNG: our validator in a worker thread (zlib
releases the GIL); JPEG: a strict libjpeg decoder in a child process."""
if info.content_type == PNG:
await asyncio.wait_for(asyncio.to_thread(verify_png, data), timeout)
else:
await verify_jpeg(data, timeout, jpeg_max_memory_mb)
def _png_segments(width: int, height: int, depth: int, color: int, interlace: int) -> list[tuple[int, int]]:
"""(byte count, row stride) of every pass of the raw (filtered) image data."""
bpp = PNG_CHANNELS[color] * depth
passes = ADAM7 if interlace else ((0, 0, 1, 1),)
segments = []
for xs, ys, dx, dy in passes:
pw = (width - xs + dx - 1) // dx if width > xs else 0
ph = (height - ys + dy - 1) // dy if height > ys else 0
if pw and ph:
stride = 1 + (pw * bpp + 7) // 8
segments.append((ph * stride, stride))
return segments
class _RawChecker:
"""Counts inflated bytes against the expected size and checks the filter byte of every row
using C-speed slicing (`piece[first::stride]`), so even 40 000 000 one-pixel rows are fast."""
VALID_FILTERS = bytes(range(5))
def __init__(self, segments: list[tuple[int, int]]):
self.segments = segments
self.index = 0 # current pass
self.offset = 0 # bytes already seen inside the current pass
self.expected = sum(size for size, _ in segments)
self.total = 0
def feed(self, piece: bytes) -> None:
self.total += len(piece)
if self.total > self.expected:
raise unsupported()
pos = 0
while pos < len(piece):
size, stride = self.segments[self.index]
take = min(size - self.offset, len(piece) - pos)
part = piece[pos:pos + take]
first = (-self.offset) % stride
if part[first::stride].translate(None, self.VALID_FILTERS):
raise unsupported()
self.offset += take
pos += take
if self.offset == size:
self.index += 1
self.offset = 0
def verify_png(data: bytes | bytearray | memoryview) -> None:
view = memoryview(data)
if bytes(view[:8]) != PNG_SIGNATURE:
raise unsupported()
pos = 8
n = len(view)
checker: _RawChecker | None = None
inflater = zlib.decompressobj()
color = None
seen_plte = seen_idat = idat_done = False
first = True
while True:
if pos + 12 > n:
raise unsupported() # truncated: no room for another chunk header + CRC
length, ctype = struct.unpack(">I4s", view[pos:pos + 8])
if length > n - pos - 12:
raise unsupported()
body = view[pos + 8:pos + 8 + length]
crc = struct.unpack(">I", view[pos + 8 + length:pos + 12 + length])[0]
if zlib.crc32(body, zlib.crc32(ctype)) != crc:
raise unsupported()
pos += 12 + length
if not all(65 <= c <= 90 or 97 <= c <= 122 for c in ctype):
raise unsupported()
if first:
if ctype != b"IHDR" or length != 13:
raise unsupported()
width, height, depth, color, compression, filtering, interlace = struct.unpack(">IIBBBBB", body)
if (not (1 <= width <= MAX_DIMENSION and 1 <= height <= MAX_DIMENSION) or color not in PNG_DEPTHS
or depth not in PNG_DEPTHS[color] or compression or filtering or interlace > 1):
raise unsupported()
checker = _RawChecker(_png_segments(width, height, depth, color, interlace))
first = False
continue
assert checker is not None
if ctype == b"IHDR":
raise unsupported()
if ctype == b"PLTE":
if seen_plte or seen_idat or color in (0, 4) or length % 3 or not 3 <= length <= 768:
raise unsupported()
seen_plte = True
elif ctype == b"IDAT":
if idat_done:
raise unsupported() # IDAT chunks must be consecutive
if color == 3 and not seen_plte:
raise unsupported()
seen_idat = True
_inflate(inflater, body, checker)
elif ctype == b"IEND":
if length != 0 or not seen_idat:
raise unsupported()
if pos != n:
raise unsupported() # bytes after IEND (polyglot files) are not accepted
break
elif not 97 <= ctype[0] <= 122:
raise unsupported() # unknown CRITICAL chunk: a decoder could not show the image
if ctype != b"IDAT" and seen_idat:
idat_done = True
# The zlib stream must end exactly with the last pixel row, without trailing data.
if not inflater.eof or inflater.unused_data or checker.total != checker.expected:
raise unsupported()
def _inflate(inflater, body: memoryview, checker: _RawChecker) -> None:
if inflater.eof:
if len(body):
raise unsupported() # data after the end of the zlib stream
return
pending: bytes | memoryview = body
try:
while pending:
piece = inflater.decompress(pending, INFLATE_STEP)
checker.feed(piece)
pending = inflater.unconsumed_tail
if inflater.eof:
if pending or inflater.unused_data:
raise unsupported()
break
except zlib.error:
raise unsupported() from None
async def verify_jpeg(data: bytes | bytearray | memoryview, timeout: float, max_memory_mb: int) -> None:
"""Decode the WHOLE file with the reference decoder (libjpeg-turbo `djpeg`) in strict mode.
`-strict` turns every libjpeg warning into a failure: "premature end of data segment"
(scan data missing or cut before EOI), "premature end of JPEG file", corrupt Huffman data.
Pillow cannot be used for this: it ignores libjpeg warnings and returns a grey picture for
a JPEG whose scan data is missing. `-scale 1/8` decodes every block but writes a 64x
smaller raster to /dev/null; `-maxmemory` caps libjpeg's working memory (coefficient
buffers of progressive files); the separate process is killed at the deadline.
"""
process = await asyncio.create_subprocess_exec(
DJPEG, "-strict", "-scale", "1/8", "-maxmemory", str(max_memory_mb * 1024), "-outfile", os.devnull,
stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.DEVNULL, stderr=asyncio.subprocess.PIPE)
stderr = b""
try:
_, stderr = await asyncio.wait_for(process.communicate(bytes(data)), timeout)
except (asyncio.TimeoutError, asyncio.CancelledError):
process.kill()
await process.wait()
raise
except (BrokenPipeError, ConnectionResetError):
await process.wait() # djpeg stopped reading: the file is already known to be broken
if process.returncode != 0:
if b"Insufficient memory" in (stderr or b""):
# A valid but extremely memory-hungry file (progressive, 4:4:4, ~40 Mpx): a limit, not corruption.
raise ImageError(413, "Изображение слишком велико для обработки")
raise unsupported()
+71 -25
View File
@@ -62,6 +62,7 @@ class Message:
sender_id: str
text: str
created_at: dt.datetime
attachment_ids: tuple[str, ...] = ()
@property
def key(self) -> SortKey:
@@ -69,7 +70,7 @@ class Message:
def to_json(self) -> dict:
return {"id": self.id, "chat_id": self.chat_id, "sender_id": self.sender_id, "text": self.text,
"created_at": rfc3339(self.created_at)}
"created_at": rfc3339(self.created_at), "attachment_ids": list(self.attachment_ids)}
def _user(row) -> User:
@@ -205,7 +206,7 @@ class PostResult:
async def _insert_message(conn: psycopg.AsyncConnection, message_id: uuid.UUID, user_id: str, chat_id: str,
text: str) -> dt.datetime:
text: str, attachment_ids: list[str]) -> dt.datetime:
"""Insert with a per-chat strictly increasing timestamp.
`UPDATE chats ... RETURNING` locks the chat row until COMMIT, so writers of one
@@ -220,9 +221,39 @@ async def _insert_message(conn: psycopg.AsyncConnection, message_id: uuid.UUID,
await conn.execute(
"INSERT INTO messages (id, chat_id, sender_id, text, created_at) VALUES (%s, %s, %s, %s, %s)",
(message_id, chat_id, user_id, _b(text), created_at))
if attachment_ids:
await conn.execute(
"INSERT INTO message_attachments (message_id, position, attachment_id) "
"SELECT %s, a.ord - 1, a.id FROM unnest(%s::uuid[]) WITH ORDINALITY AS a(id, ord)",
(message_id, attachment_ids))
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.
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' "
"FOR SHARE) a")
async def _check_attachments(conn: psycopg.AsyncConnection, chat_id: str, attachment_ids: list[str]) -> str:
if not attachment_ids:
return "ok"
cur = await conn.execute(CHECK_ATTACHMENTS, (attachment_ids, chat_id))
found, usable = await cur.fetchone()
if found != len(attachment_ids):
return "attachment_not_found"
return "ok" if usable == found else "attachment_conflict"
def _attachment_error(status: str) -> None:
if status == "attachment_not_found":
raise NotFound("Вложение не найдено")
if status == "attachment_conflict":
raise Conflict("Вложение недоступно и не может быть прикреплено")
# Claim the key. A row of a concurrent transaction that is not committed yet makes this
# INSERT wait on the unique index until that transaction ends: after COMMIT it is a conflict
# (the key is taken), after ROLLBACK the INSERT goes through. A key older than the TTL is
@@ -235,60 +266,72 @@ CLAIM_KEY = (
"WHERE idempotency_keys.created_at < clock_timestamp() - make_interval(secs => %s) "
"RETURNING 1")
STORED_RESULT = (
"SELECT k.request_hash, m.id, m.text, m.created_at FROM idempotency_keys k "
"SELECT k.request_hash, m.id, m.text, m.created_at, "
"ARRAY(SELECT ma.attachment_id FROM message_attachments ma WHERE ma.message_id = m.id ORDER BY ma.position) "
"FROM idempotency_keys k "
"JOIN messages m ON m.id = k.message_id "
"WHERE k.user_id = %s AND k.chat_id = %s AND k.key = %s")
async def post_message(db: Database, user_id: str, chat_id: str, key: str, request_hash: bytes, text: str,
ttl_seconds: int) -> PostResult:
ttl_seconds: int, attachment_ids: list[str] = ()) -> PostResult:
"""Idempotent send: the key, the message and the stored result commit in ONE transaction.
* first request -> key row + message + chat clock, all or nothing (no "pending" key is
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.
* 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).
"""
message_id = uuid.uuid4()
attachment_ids = list(attachment_ids)
async def work(conn: psycopg.AsyncConnection) -> PostResult:
await _check_member(conn, user_id, chat_id)
cur = await conn.execute(CLAIM_KEY, (user_id, chat_id, key, request_hash, message_id, ttl_seconds))
if await cur.fetchone() is not None:
created_at = await _insert_message(conn, message_id, user_id, chat_id, text)
return PostResult(Message(str(message_id), chat_id, user_id, text, created_at), False)
check = await _check_attachments(conn, chat_id, attachment_ids)
if check == "ok":
cur = await conn.execute(CLAIM_KEY, (user_id, chat_id, key, request_hash, message_id, ttl_seconds))
if await cur.fetchone() is not None:
created_at = await _insert_message(conn, message_id, user_id, chat_id, text, attachment_ids)
return PostResult(Message(str(message_id), chat_id, user_id, text, created_at,
tuple(attachment_ids)), False)
cur = await conn.execute(STORED_RESULT, (user_id, chat_id, key))
row = await cur.fetchone()
if row is None:
_attachment_error(check)
# The key row vanished between the two statements (expired-key cleanup): retryable.
raise Unavailable("Повторите запрос с тем же Idempotency-Key")
if bytes(row[0]) != request_hash:
raise Conflict("Idempotency-Key уже использован с другим сообщением")
return PostResult(Message(str(row[1]), chat_id, user_id, _s(row[2]), row[3]), True)
return PostResult(Message(str(row[1]), chat_id, user_id, _s(row[2]), row[3],
tuple(str(a) for a in row[4])), True)
return await db.run(work)
POST_MESSAGE_CALL = "SELECT o_status, o_id, o_text, o_created_at FROM post_message(%s, %s, %s, %s, %s, %s, %s)"
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[])")
async def post_message_single_call(db: Database, user_id: str, chat_id: str, key: str, request_hash: bytes,
text: str, ttl_seconds: int) -> PostResult:
"""Same contract as post_message, ONE round trip: SQL function post_message (migration 0003)
in autocommit mode, i.e. the call is the whole transaction. Track 1 improvement."""
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 =
0003 + attachment_ids) in autocommit mode, i.e. the call is the whole transaction."""
message_id = uuid.uuid4()
async def work(conn: psycopg.AsyncConnection) -> tuple:
cur = await conn.execute(POST_MESSAGE_CALL,
(user_id, chat_id, key, request_hash, message_id, _b(text), ttl_seconds))
cur = await conn.execute(POST_MESSAGE_CALL, (user_id, chat_id, key, request_hash, message_id, _b(text),
ttl_seconds, list(attachment_ids)))
return await cur.fetchone()
status, stored_id, stored_text, created_at = await db.run(work, transaction=False)
status, stored_id, stored_text, created_at, stored_attachments = await db.run(work, transaction=False)
if status == "not_found":
raise NotFound("Чат не найден")
_attachment_error(status)
if status == "conflict":
raise Conflict("Idempotency-Key уже использован с другим сообщением")
if status == "retry":
raise Unavailable("Повторите запрос с тем же Idempotency-Key")
return PostResult(Message(str(stored_id), chat_id, user_id, _s(stored_text), created_at), status == "replayed")
return PostResult(Message(str(stored_id), chat_id, user_id, _s(stored_text), created_at,
tuple(str(a) for a in stored_attachments)), status == "replayed")
async def purge_idempotency_keys(db: Database, ttl_seconds: int, batch: int = 5000) -> int:
@@ -308,12 +351,14 @@ async def purge_idempotency_keys(db: Database, ttl_seconds: int, batch: int = 50
# Two statements instead of one with "(%s IS NULL OR ...)": with a generic plan the OR
# would push the cursor into a Filter and late pages would scan the whole prefix.
HISTORY_FIRST = (
"SELECT id, sender_id, text, created_at FROM messages "
"WHERE chat_id = %s ORDER BY created_at, id LIMIT %s")
HISTORY_AFTER = (
"SELECT id, sender_id, text, created_at FROM messages "
"WHERE chat_id = %s AND (created_at, id) > (%s, %s) ORDER BY created_at, id LIMIT %s")
# attachment_ids: a correlated sub-select per returned row (at most limit + 1 = 101 lookups on the
# primary key of message_attachments), so the keyset range scan itself stays unchanged.
HISTORY_COLUMNS = ("SELECT m.id, m.sender_id, m.text, m.created_at, "
"ARRAY(SELECT ma.attachment_id FROM message_attachments ma "
"WHERE ma.message_id = m.id ORDER BY ma.position) FROM messages m ")
HISTORY_FIRST = HISTORY_COLUMNS + "WHERE m.chat_id = %s ORDER BY m.created_at, m.id LIMIT %s"
HISTORY_AFTER = (HISTORY_COLUMNS +
"WHERE m.chat_id = %s AND (m.created_at, m.id) > (%s, %s) ORDER BY m.created_at, m.id LIMIT %s")
async def list_messages(db: Database, chat_id: str, after: SortKey | None, limit: int) -> tuple[list[Message], bool]:
@@ -324,5 +369,6 @@ async def list_messages(db: Database, chat_id: str, after: SortKey | None, limit
cur = await conn.execute(HISTORY_AFTER, (chat_id, from_us(after[0]), after[1], limit + 1))
return await cur.fetchall()
rows = await db.run(work)
items = [Message(str(r[0]), chat_id, str(r[1]), _s(r[2]), r[3]) for r in rows[:limit]]
items = [Message(str(r[0]), chat_id, str(r[1]), _s(r[2]), r[3], tuple(str(a) for a in r[4]))
for r in rows[:limit]]
return items, len(rows) > limit
+151
View File
@@ -0,0 +1,151 @@
"""One-shot S3 bootstrap: `python -m messenger.s3_bootstrap` (compose service `s3-init`).
Idempotent, safe to run on every `docker compose up`:
1. waits until the S3 endpoint answers;
2. creates the bucket with the ADMIN key pair if it does not exist yet (an existing bucket and
its objects are left untouched);
3. writes the sentinel object `health/sentinel` that the API's readiness probe HEADs;
4. proves the access model and FAILS the stand (exit 1) if it is wrong:
* anonymous (unsigned) GET of an object and anonymous LIST of the bucket are refused;
* the API key pair can read/write objects of ITS bucket, but can neither list it, nor create
a bucket, nor see or read another bucket (a canary bucket created here for this check);
* the reconciler key pair can list the bucket but cannot create one.
Only file names and verdicts are printed, never key values.
"""
from __future__ import annotations
import asyncio
import os
import sys
import time
import urllib.error
import urllib.request
from botocore.exceptions import ClientError
from .config import ConfigError, Settings, _secret, s3_settings
from .errors import Unavailable
from .storage import SENTINEL_KEY, Storage
def say(text: str) -> None:
print(f"s3-bootstrap: {text}", flush=True)
def anonymous_status(url: str) -> int:
"""HTTP status of an UNSIGNED request (what anyone on the network could do)."""
try:
with urllib.request.urlopen(url, timeout=3) as response:
return response.status
except urllib.error.HTTPError as exc:
return exc.code
async def error_code(coro) -> str:
"""S3 error code of a call that is expected to be refused ('' if it succeeded)."""
try:
await coro
return ""
except ClientError as exc:
return str(exc.response.get("Error", {}).get("Code", "error"))
def settings_with(keys: tuple[str, str]) -> Settings:
base = s3_settings(credentials_required=False)
base.update(s3_access_key_id=keys[0], s3_secret_access_key=keys[1])
return Settings(host="", port=0, env="lab", log_level="info", shutdown_timeout=0, max_body_bytes=1024, **base)
async def run() -> int:
admin = (_secret("S3_ADMIN_ACCESS_KEY_ID", True), _secret("S3_ADMIN_SECRET_ACCESS_KEY", True))
api = (_secret("S3_ACCESS_KEY_ID", True), _secret("S3_SECRET_ACCESS_KEY", True))
reconciler = (_secret("S3_RECONCILER_ACCESS_KEY_ID", True), _secret("S3_RECONCILER_SECRET_ACCESS_KEY", True))
settings = settings_with(admin)
bucket = settings.s3_bucket
admin_s3, api_s3, rec_s3 = Storage(settings), Storage(settings_with(api)), Storage(settings_with(reconciler))
for s in (admin_s3, api_s3, rec_s3):
await s.open()
try:
deadline = time.monotonic() + 60
while True:
try:
await admin_s3._call("list_buckets", admin_s3.client.list_buckets())
break
except Unavailable:
if time.monotonic() > deadline:
say("S3 endpoint does not answer within 60 s")
return 1
await asyncio.sleep(1)
code = await error_code(admin_s3.client.head_bucket(Bucket=bucket))
if code == "":
say(f"bucket {bucket}: exists, kept as is")
elif code in ("404", "NoSuchBucket", "NotFound"):
code = await error_code(admin_s3.client.create_bucket(Bucket=bucket))
if code not in ("", "BucketAlreadyOwnedByYou", "BucketAlreadyExists"):
say(f"bucket {bucket}: create failed ({code})")
return 1
say(f"bucket {bucket}: created")
else:
say(f"bucket {bucket}: head failed ({code})")
return 1
await admin_s3.put(SENTINEL_KEY, b"ok\n", "text/plain", {})
say(f"sentinel {SENTINEL_KEY}: written")
canary = bucket + "-canary"
if await error_code(admin_s3.client.head_bucket(Bucket=canary)):
await admin_s3.client.create_bucket(Bucket=canary)
await admin_s3.client.put_object(Bucket=canary, Key="canary", Body=b"secret of another bucket")
failures = []
base = settings.s3_endpoint.rstrip("/")
for what, url in (("anonymous GET object", f"{base}/{bucket}/{SENTINEL_KEY}"),
("anonymous LIST bucket", f"{base}/{bucket}?list-type=2")):
status = anonymous_status(url)
verdict = "refused" if status in (401, 403) else "ALLOWED"
say(f"{what}: HTTP {status} -> {verdict}")
if verdict != "refused":
failures.append(what)
probe = "health/bootstrap-probe"
checks = [
("api: read object", await error_code(api_s3.client.head_object(Bucket=bucket, Key=SENTINEL_KEY)), True),
("api: write object", await error_code(api_s3.client.put_object(Bucket=bucket, Key=probe, Body=b"p")), True),
("api: delete object", await error_code(api_s3.client.delete_object(Bucket=bucket, Key=probe)), True),
("api: list bucket", await error_code(api_s3.client.list_objects_v2(Bucket=bucket, MaxKeys=1)), False),
("api: create bucket", await error_code(api_s3.client.create_bucket(Bucket=bucket + "-x")), False),
("api: read object of another bucket",
await error_code(api_s3.client.head_object(Bucket=canary, Key="canary")), False),
("reconciler: list bucket", await error_code(rec_s3.client.list_objects_v2(Bucket=bucket, MaxKeys=1)), True),
("reconciler: create bucket", await error_code(rec_s3.client.create_bucket(Bucket=bucket + "-x")), False),
]
for what, code, allowed in checks:
ok = (code == "") == allowed
say(f"{what}: {'allowed' if code == '' else 'refused (' + code + ')'} -> {'OK' if ok else 'WRONG'}")
if not ok:
failures.append(what)
visible = [b["Name"] for b in (await api_s3.client.list_buckets()).get("Buckets", [])]
say(f"api: buckets visible in ListBuckets: {visible} -> {'OK' if canary not in visible else 'WRONG'}")
if canary in visible:
failures.append("api sees another bucket")
if failures:
say("access model is WRONG: " + ", ".join(failures))
return 1
say("done: private bucket, least-privilege keys verified")
return 0
finally:
for s in (admin_s3, api_s3, rec_s3):
await s.close()
def main() -> int:
try:
return asyncio.run(run())
except ConfigError as exc:
print(f"configuration error: {exc}", file=sys.stderr)
return 2
if __name__ == "__main__":
os.environ.setdefault("PYTHONUNBUFFERED", "1")
sys.exit(main())
+173
View File
@@ -0,0 +1,173 @@
"""Private S3 bucket (SeaweedFS in compose) through aiobotocore, with a hard deadline per call.
* Credentials: the API's OWN key pair (S3_ACCESS_KEY_ID_FILE / S3_SECRET_ACCESS_KEY_FILE ->
Docker secrets). SeaweedFS grants it only Read + Write on this one bucket: no List, no
Admin (cannot create/delete buckets or read other buckets). The admin pair is given only
to the S3 server and the one-shot bootstrap (infra/seaweedfs, messenger.s3_bootstrap).
* Deadlines: botocore connect/read timeouts AND an asyncio deadline around every call, so a
paused or unreachable S3 costs at most S3_TIMEOUT_SECONDS, then 503 `unavailable`. Unlike a
libpq query (db.py), an aiohttp request can be cancelled safely, so the call is cancelled.
* No retries inside botocore (max_attempts=1): the request handler decides; an upload is not
retried behind the client's back (contract: a retry may create another object).
* Checksums: Content-MD5 is sent with every PUT, so S3 rejects bytes damaged on the way.
botocore's newer default (aws-chunked CRC32 trailers) is switched off ("when_required"):
plain, widely supported S3 requests.
"""
from __future__ import annotations
import asyncio
import base64
import contextlib
import hashlib
import logging
from collections.abc import AsyncIterator
from dataclasses import dataclass
from aiobotocore.config import AioConfig
from aiobotocore.session import get_session
from botocore.exceptions import BotoCoreError, ClientError
from .config import Settings
from .errors import Unavailable
log = logging.getLogger("messenger.storage")
MISSING_CODES = {"NoSuchKey", "404", "NotFound"}
class ObjectMissing(Exception):
"""The key does not exist in the bucket (definitive answer of S3, not an outage)."""
@dataclass(frozen=True, slots=True)
class ObjectHead:
size: int
content_type: str
metadata: dict[str, str]
@dataclass(frozen=True, slots=True)
class ListedObject:
key: str
size: int
last_modified_ts: float # POSIX seconds
class Storage:
def __init__(self, settings: Settings, *, access_key: str | None = None, secret_key: str | None = None):
self.bucket = settings.s3_bucket
self.timeout = settings.s3_timeout_seconds
self._endpoint = settings.s3_endpoint
self._region = settings.s3_region
self._access_key = access_key or settings.s3_access_key_id
self._secret_key = secret_key or settings.s3_secret_access_key
self._config = AioConfig(
connect_timeout=settings.s3_connect_timeout_seconds,
read_timeout=settings.s3_timeout_seconds,
retries={"total_max_attempts": 1, "mode": "standard"},
s3={"addressing_style": "path"},
signature_version="s3v4",
request_checksum_calculation="when_required",
response_checksum_validation="when_required",
max_pool_connections=settings.s3_max_connections,
)
self._stack: contextlib.AsyncExitStack | None = None
self._client = None
async def open(self) -> None:
self._stack = contextlib.AsyncExitStack()
session = get_session()
self._client = await self._stack.enter_async_context(session.create_client(
"s3", endpoint_url=self._endpoint, region_name=self._region,
aws_access_key_id=self._access_key, aws_secret_access_key=self._secret_key, config=self._config))
async def close(self) -> None:
if self._stack is not None:
await self._stack.aclose()
self._stack = None
@property
def client(self):
assert self._client is not None, "Storage.open() was not called"
return self._client
async def _call(self, what: str, coro, deadline: float | None = None):
"""Await one S3 call under the deadline; outages become Unavailable (503)."""
try:
return await asyncio.wait_for(coro, deadline or self.timeout)
except ClientError as exc:
code = str(exc.response.get("Error", {}).get("Code", ""))
if code in MISSING_CODES:
raise ObjectMissing() from None
log.warning("s3_error", extra={"fields": {"op": what, "code": code}})
raise Unavailable("Хранилище файлов временно недоступно") from None
except (asyncio.TimeoutError, BotoCoreError, OSError) as exc:
log.warning("s3_unavailable", extra={"fields": {"op": what, "error": type(exc).__name__}})
raise Unavailable("Хранилище файлов временно недоступно") from None
async def put(self, key: str, data: bytes | bytearray, content_type: str, metadata: dict[str, str]) -> None:
md5 = base64.b64encode(hashlib.md5(data, usedforsecurity=False).digest()).decode("ascii")
# Time budget grows with the size: 10 MiB over a local network takes well under a second.
deadline = self.timeout + len(data) / (8 * 1024 * 1024)
await self._call("put", self.client.put_object(
Bucket=self.bucket, Key=key, Body=bytes(data), ContentType=content_type, ContentLength=len(data),
ContentMD5=md5, Metadata=metadata), deadline)
async def head(self, key: str) -> ObjectHead:
response = await self._call("head", self.client.head_object(Bucket=self.bucket, Key=key))
return ObjectHead(int(response["ContentLength"]), response.get("ContentType", ""),
dict(response.get("Metadata", {})))
async def open_stream(self, key: str) -> tuple[int, AsyncIterator[bytes]]:
"""Start a GET: (Content-Length, async iterator of chunks). The headers arrive under the
deadline; every chunk read afterwards has its own deadline (a stalled S3 aborts the
download instead of hanging the connection)."""
response = await self._call("get", self.client.get_object(Bucket=self.bucket, Key=key))
body = response["Body"]
size = int(response["ContentLength"])
async def chunks() -> AsyncIterator[bytes]:
try:
while True:
chunk = await self._call("get_body", body.read(64 * 1024))
if not chunk:
return
yield chunk
finally:
body.close()
return size, chunks()
async def read_all(self, key: str, limit: int) -> bytes:
"""Whole object (reconciler checksum verification); refuses objects above `limit`."""
size, chunks = await self.open_stream(key)
if size > limit:
raise Unavailable("Объект больше допустимого размера")
parts = [chunk async for chunk in chunks]
return b"".join(parts)
async def delete(self, key: str) -> None:
"""Idempotent: deleting a missing key is a success in S3."""
await self._call("delete", self.client.delete_object(Bucket=self.bucket, Key=key))
async def list_page(self, prefix: str, start_after: str | None, page_size: int) -> tuple[list[ListedObject], bool]:
"""One ListObjectsV2 page in key order (needs the List permission: reconciler only)."""
params = {"Bucket": self.bucket, "Prefix": prefix, "MaxKeys": page_size}
if start_after:
params["StartAfter"] = start_after
response = await self._call("list", self.client.list_objects_v2(**params))
items = [ListedObject(o["Key"], int(o["Size"]), o["LastModified"].timestamp())
for o in response.get("Contents", [])]
return items, bool(response.get("IsTruncated"))
async def ping(self, timeout: float = 1.0) -> bool:
"""Readiness: HEAD of the sentinel object written by the bootstrap. Proves that S3
answers, the bucket exists and the API credentials are accepted (Read permission)."""
try:
await self._call("ping", self.client.head_object(Bucket=self.bucket, Key=SENTINEL_KEY), timeout)
return True
except (Unavailable, ObjectMissing):
return False
SENTINEL_KEY = "health/sentinel"
OBJECT_PREFIX = "o/" # every attachment original lives under o/<random 128-bit hex>
+56 -4
View File
@@ -8,7 +8,9 @@ from __future__ import annotations
import json
import re
import unicodedata
from typing import Any
from urllib.parse import parse_qsl
from .errors import InvalidRequest
@@ -58,7 +60,7 @@ def check_fields(data: dict[str, Any], required: set[str], optional: frozenset[s
raise InvalidRequest("Отсутствуют обязательные поля: " + ", ".join(sorted(missing)))
def text_field(data: dict[str, Any], name: str, max_len: int, *, min_len: int = 1) -> str:
def text_field(data: dict[str, Any], name: str, max_len: int, *, min_len: int = 1, allow_blank: bool = False) -> str:
value = data[name]
if not isinstance(value, str):
raise InvalidRequest(f"Поле {name} должно быть строкой")
@@ -68,7 +70,7 @@ def text_field(data: dict[str, Any], name: str, max_len: int, *, min_len: int =
# len() of a Python str is the number of Unicode code points, as the contract requires.
if not min_len <= len(value) <= max_len:
raise InvalidRequest(f"Длина поля {name} должна быть от {min_len} до {max_len} символов")
if value.strip() == "":
if value.strip() == "" and not allow_blank:
raise InvalidRequest(f"Поле {name} не может состоять только из пробелов")
return value
@@ -113,8 +115,8 @@ def member_ids_field(data: dict[str, Any]) -> list[str]:
def attachment_ids_field(data: dict[str, Any]) -> list[str]:
"""Optional `attachment_ids` (lab 4+). Absent and [] are the same message (contracts/README.md);
lab 3 has no attachments yet, so a non-empty list can only reference unknown objects."""
"""Optional `attachment_ids` (lab 4+): absent and [] are the same message (contracts/README.md);
null or any other non-array type is 400."""
value = data.get("attachment_ids", [])
if not isinstance(value, list):
raise InvalidRequest("attachment_ids должен быть массивом")
@@ -145,3 +147,53 @@ def parse_limit(raw: str | None) -> int:
if LIMIT_RE.fullmatch(raw) is None or not 1 <= int(raw) <= MAX_LIMIT:
raise InvalidRequest(f"limit должен быть целым числом от 1 до {MAX_LIMIT}")
return int(raw)
# Lab 4: the display name of an attachment. Never used as a path or an S3 key, but it is shown
# to other users, so characters that change how a name LOOKS are refused: control characters
# (C0, DEL, C1 such as U+0085), line/paragraph separators, unpaired surrogates and bidi controls
# ("photo<U+202E>gnp.exe" would display as "photoexe.png").
BIDI_CONTROLS = frozenset("\u061c\u200e\u200f\u202a\u202b\u202c\u202d\u202e\u2066\u2067\u2068\u2069")
FORBIDDEN_CATEGORIES = frozenset({"Cc", "Cs", "Zl", "Zp"})
def query_items(raw: bytes) -> list[tuple[str, str]]:
"""Strict query parsing: percent-encoded UTF-8 only (Starlette would replace bad bytes by U+FFFD)."""
try:
return parse_qsl(raw.decode("ascii"), keep_blank_values=True, encoding="utf-8", errors="strict")
except (UnicodeDecodeError, ValueError):
raise InvalidRequest("Параметры запроса должны быть в UTF-8 с percent-encoding") from None
def filename_param(query: list[tuple[str, str]]) -> str:
value = single_query_param(query, "filename")
if value is None:
raise InvalidRequest("Параметр filename обязателен")
if not 1 <= len(value) <= 255:
raise InvalidRequest("Длина filename должна быть от 1 до 255 символов")
if value.strip() == "" or value in {".", ".."}:
raise InvalidRequest("filename не может состоять только из пробелов или точек")
if "/" in value or "\\" in value:
raise InvalidRequest("filename не может содержать / или \\")
for char in value:
if unicodedata.category(char) in FORBIDDEN_CATEGORIES or char in BIDI_CONTROLS:
raise InvalidRequest("filename содержит управляющие символы")
return value
ATTACHMENT_VARIANTS = frozenset({"thumbnail", "crop", "optimized"}) # exist in the schema, not in lab 4
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)."""
value = single_query_param(query, "variant")
if value is None or value == "original":
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."""