# 9202, 2026-10-22 14:35–14:40 UTC — after the backup without the model cache
- POST /api/stacks/immich/stop → 200; `docker volume rm immich_immich_ml_cache` (as on a rebuilt box: no cache).
- The local unit restore (POST /backup/restore stack_name=immich snapshot_id=helyi) is REFUSED by design (R-538): Immich's
  unit holds no photo files, so the database is not put back over files it does not hold — the route for Immich is the
  second drive's or the off-site whole restore. 9202 has neither target, so the restore leg itself was NOT run here.
  What changed is only the cache volume's absence from the unit, so the proof is the state ANY such restore leaves:
- POST /api/stacks/immich/start → 200; all four containers healthy; the cache volume holds 4096 B (empty).
- Smart search's call, inside immich-machine-learning: POST localhost:3003/predict {clip.textual ViT-B-32__openai}
  → HTTP 200, embedding length 512, 9.2 s (the model fetched); again → 200, 512, 0.1 s. The cache now 610 978 474 B.
