A large import starts a chain of background work. An original can be stored correctly while its thumbnail, search index or face results are still being prepared. Diagnose the stage that is behind before restarting the whole stack or rebuilding every result.
Start with one affected item
Choose an item you can identify without exposing private material. Record when it was imported and what fails: opening the original, seeing a preview, playing a video, finding it in search or seeing a recognized person.
Try the original first. If it does not open, investigate storage and permissions before machine learning. A successful thumbnail is not proof that the original is still accessible.
Follow the processing chain
New uploads trigger metadata, storage organization where enabled, and thumbnails. Search indexing, face work, text recognition, video copies and enrichment follow. Work downstream can wait for upstream results.
Open Settings → Computers & background work → Background work. Read the summary and Now list. Use Show all kinds of work to locate a missing queue, then Manage to inspect its waiting, active and failed jobs.
| Symptom | First place to look |
|---|---|
| Wrong or absent capture details | Reading photo details and the source’s embedded metadata |
| Blank or missing previews | Making thumbnails, file access and free disk space |
| A photo does not appear in smart search | Preparing photos for search and the machine-learning service |
| Faces have not formed useful people | Finding faces, then Recognising people |
| Original video downloads but browser playback fails | Converting videos, playback copies and client compatibility |
| Descriptions are absent | Writing captions and its configured model |
Decide whether work is progressing
Compare completed/remaining counts over a useful interval and inspect the error from a failed job. A queue with active work and falling totals is different from a queue that repeatedly fails on the same item.
Record the hardware, storage location, queue and elapsed time if you want advice on performance. An item count by itself is not a useful benchmark: video lengths, RAW formats, model choices and storage latency all affect the work.
Fix a cause, then retry narrowly
- Unreadable path: reconnect the volume and check access from the relevant worker. Do not clear library records to make an error disappear.
- Disk full: free appropriate space or extend storage. Keep original media and recovery files intact.
- Model download failure: check connectivity to the configured source and model-cache space.
- Out of memory: reduce competing work, use a supported model suited to available memory, and test one item.
- Worker cannot find media: compare its mounts with the server’s container paths.
On the Failed tab, choose Try again after correcting the cause. Catch up queues items without results. Do it all again is a deliberate whole-library rerun; use it for a reason, such as rebuilding search after changing its model.
Viewing one account in Settings does not limit a maintenance command to that account. Read the command’s impact review before applying a server-wide change.
Keep heavy tasks from competing unnecessarily
Begin with CPU processing and a small sample. Video acceleration and machine learning have different requirements. A GPU that encodes video well can still lack memory for a description model. Include the exact GPU, runtime and currently free dedicated memory when asking for model help; unified/shared memory also serves the operating system.
Compare descriptions on a sample before enabling a large backfill. Keep model settings stable until you can judge the result and its resource use. Cloud GPU is available for supported heavy tasks, with the destination and cost estimate reviewed before confirming a job.
A useful support reply
Include the server/app release, installation method, affected queue, one sanitized error, storage type, CPU/RAM/GPU and whether counts are changing. Say whether originals open. That separates storage, processing and client issues quickly.
References: Jobs and queues · Choose processing · Library Care