Removes the OpenCLIP-on-ONNX classifier and everything that fed or
consumed it:
- backend: app/services/vision/, app/tasks/vision.py,
app/services/feature_flags.py, app/routers/features.py — all
deleted; admin AI/feature-flag endpoints and the worker-vision
bootstrap call gone. Photo.needs_review and its index dropped.
- frontend: AI Settings tab, useFeaturesQuery hook, FeatureFlag
types, "Needs Review" sidebar entry + filter, needs_review filter
URL param all gone.
- infra: worker-vision compose service + models_data volume deleted;
worker-light command no longer runs bootstrap_models; the db
image switches from pgvector/pgvector:pg16 to postgres:16; backend
Dockerfile drops the dedicated torch RUN layer; requirements.txt
drops torch/torchvision/open-clip-torch/onnxruntime.
Alembic 0019_drop_ai_remnants:
- drops photos.needs_review + ix_photos_needs_review
- DROP EXTENSION IF EXISTS vector (must run before the image swap;
the new postgres:16 doesn't ship pgvector)
New scripts/full_refresh.py: one-shot DB ↔ filesystem reconciliation.
Runs cleanup_data_integrity, scans every active SourceRoot inline
(no celery dependency so the worker can be stopped), hard-prunes
photo + folder rows for files that are gone, removes orphan
/data/thumbs/{user}/{photo}/ directories. New helper
prune_orphan_thumbnails in cleanup.py.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Long videos blocked /playback for the entire encode duration. The fix
is to populate the cache before the user clicks, not when they click.
Changes:
- Extract ffprobe + ffmpeg helpers to services/video.py so the request
handler and the background task share one sync implementation. The
endpoint wraps calls in asyncio.to_thread; celery just calls them.
- New tasks/video.py with pretranscode_video. Idempotent: skips when
the cache is already current and skips passthrough-safe sources
(h264 in mp4/m4v/webm). 30-min task time limit so the long-tail
files (3GP archive, multi-minute 1080p clips) still complete.
- scan_folder now dispatches pretranscode_video alongside
generate_thumbnails / extract_metadata for any new video row.
- POST /library/maintenance/backfill-video-cache enqueues every
active video so the existing library catches up.
- libx264 preset bumped from fast to veryfast. ~2x throughput on this
CPU-only box, output a few % larger but well within disk budget.
- /playback simplifies to: cache check, passthrough if h264 in
web-safe container, else sync transcode (still there as fallback
for races against the queued task).
Once the backfill task drains, /playback should be near-instant for
every video. Any video added afterwards is pre-transcoded at scan
time, so the user keeps that property going forward.
End-to-end webhook flow is proven on this NC instance (NodeCreated +
NodeWritten both fired and dispatched scan_folder on a PUT test), so
the watchfiles-based polling layer is no longer needed.
- scanner.start_initial_scan no longer queues watch_folders on boot.
- scan.watch_folders kept as a one-line no-op shim so any leftover
apply_async in flight from the previous deploy doesn't crash a
worker. Will be deleted entirely after the queue drains.
- celery.py reroutes watch_folders to the `default` queue (worker-light)
so the no-op shim actually completes — the `watcher` queue is dead.
- docker-compose drops the mulita-worker-watcher service. Its celery
--beat responsibility (firing discard_missing_photos_beat every 30
min) moves to worker-light's command.
Latency note: NC dispatches webhook events through its background-job
queue, currently run by cron */5. After this commit lands you'll want
to tighten cron to */1 so new uploads land in mule within ~60s instead
of up to 5 min.
Two related fixes for the Nextcloud library lifecycle.
1. DELETE /api/v1/nextcloud/source-roots/{id} now actually deletes
the SourceRoot, every Folder under it, and every Photo in those
folders (Nextcloud files untouched). Was a soft-deactivate
(is_active=false) that left the rows around forever, so re-adding
the same path resurrected ghosts and prune-missing reported zero.
Returns {deleted_photos, deleted_folders}; the Settings UI toasts
the count and invalidates photos/folders/stats so cached lists
don't show ghosts. photo_tags and heap_photos already cascade via
ON DELETE CASCADE; FolderShare uses a stringly-typed folder_id
with no FK so cleaned up explicitly.
2. The watcher (watch_folders task) was getting killed every five
minutes by the global task_soft_time_limit=300 in app/tasks/celery.py
despite passing soft_time_limit=None on the decorator (None falls
back to the worker default in this Celery version). Override with
soft_time_limit=0, time_limit=0 (= unlimited) so the watch loop
actually stays alive. The 'Soft time limit (300s) exceeded' /
'Worker exited prematurely' lines should stop in worker-watcher
logs.
3. Added discard_missing_photos() in services/cleanup.py — a soft
variant of prune_missing_photos that walks every present source
root, checks os.path.exists for each non-discarded Photo, and
flips is_discarded=true on the missing ones (UPDATE not DELETE).
Wired as discard_missing_photos_beat in tasks/scan.py and
scheduled every 30 min via celery beat. Beat runs in-process on
worker-watcher (--beat flag in compose) — there's only ever one
watcher and we don't need a separate container.
Hard delete remains manual via prune-missing for users who want to
review before committing. The beat catch-up only soft-discards (file
gone -> mule-image trash, restorable).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drops face recognition, OCR, object detection, and semantic embeddings.
The sole remaining vision task is a CLIP-based binary classifier
(photography vs other); photos in "other" get needs_review=true so
screenshots, documents, memes and scans can be triaged from a new
filter pill in the UI.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses 16 robustness, transparency, and performance issues across
the Celery media processing pipeline:
Critical:
- Singleton DB engine in vision tasks (was leaking one per task call)
- acks_late + task_reject_on_worker_lost so crashed workers don't lose tasks
- Global soft/hard time limits (5/10 min) to prevent hung worker slots
- Thumbnail copy-before-resize (in-place mutation degraded larger sizes)
- backfill_vision now checks each task type independently (OCR, faces, etc.)
- Parameterized LIMIT in backfill_vision (was f-string SQL injection)
High:
- try/except + retry(max=3) on all vision inference tasks
- extract_metadata writes processing_error on exiftool failure
- PIL Image handles closed in _load_thumb/_load_original
- Scan progress Redis keys auto-expire after 1 hour
- Watcher lock renewal is wall-clock based (30s) not event-count based
- worker_process_init signal warms up vision models on startup
Medium:
- Explicit task_routes for every task name (wildcards never matched)
- app.services.metadata added to Celery include list
- POST /maintenance/recover-stuck endpoint for photos stuck in processing
- Docker healthchecks for worker-light, worker-vision, and Redis
- Task ID in vision log lines for distributed tracing
- Bare except:pass narrowed to specific exceptions
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Move watch_folders to dedicated 'watcher' queue with its own
single-concurrency container so it never blocks scan/thumbnail slots
- Add get_current_user_media dependency that accepts ?token= query
param for <img src> / <video src> media endpoints (thumb, original,
proxy) — fixes 401 on thumbnails
- Append JWT token to all media URLs in the frontend
- Add missing 'memories' case in sidebar navigation switch
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Wire the full embedding flow:
- Rewrite Embedding model to use pgvector Vector(512) with HNSW index
- Add embed_photo, vision_fanout, backfill_vision Celery tasks on
dedicated `vision` queue
- Hook vision_fanout into generate_thumbnails completion
- Add POST /api/v1/photos/search with hybrid RRF ranking (semantic-only
for now; FTS leg added in PR5)
- Stub ocr_photo, detect_objects, extract_faces tasks for later PRs
Migration 0003 drops/recreates the embeddings table (was never populated).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
