Serializes turbo tasks (concurrency=1): api and worker e2e suites share
one MySQL/MinIO and resetDb TRUNCATEs raced when run in parallel.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Task 21: PublishService.processPhoto streams the original from storage,
runs the sharp pipeline, puts thumb/browse variants under
photo-live/dev/albums/{albumId}/variants/{photoId}/1/, then commits ONE
Prisma transaction: photo_variants upserts (thumb+browse), a conditional
photos update WHERE processingState IN (uploaded, processing_failed) that
flips to ready+published, stamps publishedAt/width/height, clears
errorCode and bumps entityVersion, and the outbox_events row
(photo.published, payload = contracts PhotoPublishedPayload with the
post-increment entityVersion). A zero-row conditional update means a
concurrent winner already published -> idempotent 'already' with no
second event and no extra version bump. The BullMQ worker wrapper
(QUEUE_PHOTO_PROCESS, concurrency 2, jobId 'process-{photoId}' without
the reserved ':') claims photo_jobs pending->active (attempt++), marks
done for published/already/skipped alike, and rethrows pipeline/storage
errors for the producer-configured 3-attempt exponential backoff; the
final attempt (BullMQ increments attemptsMade only after
moveToFinished, so in-processor it is 0-based: current = attemptsMade+1)
marks photo_jobs failed + photo processing_failed with the
PipelineError code while visibility stays unpublished. sweepOnce()
claims due pending rows (nextRetryAt IS NULL OR < now) and re-enqueues
stale (>5min) active rows with jobId dedupe; startSweepInterval(30s)
boot-scans then loops, and main.ts wires pino logging, dedicated redis
connections and graceful SIGINT/SIGTERM shutdown.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Task 20: services/worker package (@photo/worker, commonjs) with jest
unit/e2e configs mirroring api conventions (root .env via test/load-env.ts,
moduleNameMapper to @photo/contracts/db-cn/storage TS sources, maxWorkers 1,
test/reset.ts = MySQL TRUNCATE + MinIO photo-live/dev prefix wipe) and the
sharp processing pipeline runPipeline(): metadata whitelist gate
(jpeg/png/webp/avif/heif/tiff, else PipelineError UNSUPPORTED_TYPE),
decode failures mapped to PipelineError DECODE_FAILED, no-arg .rotate()
auto-orientation (EXIF 5-8 swaps stored width/height), thumb <=480w q75
and browse <=1920 q82 JPEG variants with withoutEnlargement, and no
withMetadata so outputs carry no EXIF/GPS at all. sharp 0.35 prebuilt
loads script-free under pnpm 10; its win32 libvips lacks a HEVC encoder,
so the HEVC-HEIF unit case self-skips with a console note (G2 verifies
with real iPhone samples) while the AV1-HEIF case covers the same
libheif decode path / 'heif' format bucket.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Task 19: POST /uploads/{id}/renew re-signs the SAME staging key for pending
sessions (uploadEnabled=false -> 409 ALBUM_UPLOAD_DISABLED; non-owner/unknown/
removed-member -> 403 AUTH_FORBIDDEN, indistinguishable per contract) and
slides the session expiry — renew is the documented recovery path for an
expired credential. POST /uploads/{id}/complete verifies the staging object
out-of-transaction first (head: missing / byte-count mismatch / contentType
mismatch — the server-side backstop for presign not signing content-type),
then copyIfAbsent's it to the finals key with forbid-overwrite semantics,
then in ONE transaction updates photo(bytes/uploadedAt/uploaded), claims the
session pending->confirmed via conditional updateMany (concurrent completes:
exactly one winner creates the photo_jobs row + upload.complete audit), and
enqueues QUEUE_PHOTO_PROCESS post-commit (jobId derived from photoId for
dedupe; failure only logs — Task 21 sweep compensates on photo_jobs). Verify
failures are request-level 400s: the session stays pending and retryable.
Repeat complete on a confirmed session re-runs copyIfAbsent (created:false,
staging already cleaned is fine) and returns the current processingState —
no second job/audit/enqueue (A03). GET /uploads lists the caller's own
sessions newest-first with (createdAt,id) keyset pagination and the joined
photo processingState/visibility. Both producer and test Queue instances now
quit their caller-owned ioredis connection (BullMQ does not take ownership),
otherwise jest hangs on open handles.
Co-Authored-By: Claude Code <noreply@anthropic.com>
POST /v1/uploads (Task 18, A03): @@unique(userId, clientTaskId) idempotency
anchor — same params return the original session (201, no duplicate photo);
param drift (album/bytes/contentType/extension) -> 409 UPLOAD_SESSION_CONFLICT.
photos.objectKey is set to the FINAL originals key at init while the session
holds the staging key; IDs come from an app-side cuid-like generator so both
self-referencing keys are written in one transaction. Credential is a real
MinIO presigned PUT (SIGNED_URL_TTL_SECONDS, lazily read). Permission order:
album invisible 404 -> member 403 -> uploadEnabled 409 -> size/ext 400 ->
category ownership 400. No audit on init (audit anchors on complete).
Pre-step for Task 19: QUEUE_PHOTO_PROCESS contract constant + lazy-safe
BullMQ queue producer (Redis down at boot must not crash the app; enqueue
failure only logs — sweep compensates). e2e reset now also wipes the MinIO
photo-live/dev prefix; jest maps @photo/storage to TS source.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Task 15: POST/GET /v1/albums, GET /v1/albums/{id}.
- create: single transaction writes albums + album_translations +
album_members(owner) + audit_logs(album.create); publicToken 24-char
[0-9a-z] with one P2002 retry; duplicate-locale input rejected as 400.
- dto: class-validator per contract AlbumCreateRequest, incl. cross-field
rules (defaultLocale must be covered by translations; endAt>startAt when
present; timezone string-only per P1a scope).
- get: PermissionsService.canViewAlbum gate -> 404 ALBUM_NOT_FOUND for
non-member/nonexistent alike (anti-enumeration); admin bypass; joins
categories (empty until Task 17).
- list: admin sees all, members see joined albums; stable createdAt DESC,
id DESC keyset pagination via new shared/cursor.ts (encode/decode +
typed where-builder, designed for Task 23 photo list reuse).
- e2e: services/api/test/albums.e2e-spec.ts (12 cases, seeded sessions).
Co-Authored-By: Claude Code <noreply@anthropic.com>
Task 14 (A02 foundation): PermissionsService with fresh-DB-per-call
canViewAlbum/canManageAlbum/canUploadToAlbum/canManagePhoto over the
album_members @@unique([albumId,userId]) compound key; admin bypass,
owner/collaborator matrix per requirements doc section 2, removed
members lose access on the very next call (no caching). Boolean-only
returns; callers map 404/403. AlbumsModule shell wired into app module
(controllers arrive in Task 15). jest maxWorkers=1: DB-backed specs
share one MySQL database and reset it via TRUNCATE, so parallel spec
files wipe each other's seed data.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Spec review (live negative test) found s3-request-presigner unconditionally
unsigns content-type: header-less PUT returns 200, so the previous docblock
claiming signature enforcement was a false guarantee. Content-Type is now
documented as convention-only; head() returns contentType so the complete
step (Task 19) can verify it server-side against the session declaration.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Task 10: photos/photo_variants/upload_sessions/photo_jobs/outbox_events
on the CN schema (flat string FKs, no Prisma relations); migration
20260917120827_photo_upload_domain rewritten to utf8mb4_0900_bin per the
固化 collation recipe; reset.ts table order extended; one semantics e2e
case for photo_jobs @@unique([photoId, type]) and the upload_sessions
A03 idempotency anchor @@unique([userId, clientTaskId]).
Co-Authored-By: Claude Code <noreply@anthropic.com>
浅念