Resume Paperless with persistent broker and staging storage

apps/paperless/README.md

@@ -13,15 +13,37 @@ The `paperless-storage` Application owns the pre-existing `paperless-data` and
 recreates them. The scanner action copies scans to the `consume` subdirectory
 of `paperless-data`; it does not use email or a recipient address.
 
-PostgreSQL currently remains the existing manual Helm release `postgresql`
-at chart 1.6.8 with the PostgreSQL 17.11 image. Its Argo declaration is staged
-but intentionally not activated: a server-side dry-run of the current chart
-against the live StatefulSet reported an immutable-spec/field-ownership
-conflict. Do not delete or recreate its StatefulSet or PVC to force adoption.
-The `postgres-data-postgresql-0` PVC and `paperless-postgres` Secret must be
-preserved. Valkey remains the existing ephemeral Deployment for now; its
-upstream chart changes the Deployment selector, so its staged Application is
-also intentionally not activated. Resolve these as separate, verified cutovers.
+PostgreSQL is the dedicated CloudNativePG cluster `paperless-cnpg`, reached
+through the existing `postgresql` Service alias. Preserve its PVC and the
+retired `postgres-data-postgresql-0` rollback claim.
+
+`paperless-valkey` now owns the upstream Valkey chart and retains the `valkey`
+Service address. Its AOF is stored in `paperless-data/valkey` with every-second
+fsync. Paperless stages uploads and PDF edits in `paperless-data/scratch`, not
+an ephemeral container filesystem. Both paths belong to the existing CephFS
+backup selection. The scheduled file mirror is not an atomic queue snapshot;
+for queue recovery, quiesce Paperless and capture a consistent Valkey RDB as
+well as the staged files, database and application Secret. Do not call an
+unmounted `/mnt/backup` a successful backup destination.
+
+Paperless has no hostPath or dependency on `/mnt` or `/komposthaufen`. Its
+preference for `arschrock` is soft; a cordon permits scheduling on another node.
+Four import workers limit database connection bursts. Drain active/reserved
+Celery tasks before replacements: durable storage does not replay every task
+that was already acknowledged by a worker.
+
+### Stale task history (2026-10-28)
+
+Seventy old in-progress imports already had real documents. Original SHA-256,
+OCR content and thumbnail checks were used before reconciling their results;
+do not blindly mark pending tasks successful or re-upload duplicates. Historical
+failures were retained and acknowledged only after verifying successful retries
+or complete replacement splits. Two old bulk metadata jobs lacked their broker
+payloads and were cancelled, not falsely reported successful.
+
+For a future incident, compare task records with Celery active/reserved work,
+broker queues, staged files and existing documents before retrying. Back up
+the database and task records first; never delete the task list to hide errors.
 
 `paperless-secret-key`, `paperless-postgres`, and `paperless-oidc` remain
 cluster-managed Secrets referenced by name. They are not copied into this

apps/paperless/values.yaml

@@ -1,4 +1,4 @@
-replicaCount: 0
+replicaCount: 1
 url: https://paper.brunner.ninja
 timeZone: Europe/Vienna
 ocrLanguage: deu+eng

bootstrap/paperless-valkey.yaml

@@ -20,6 +20,6 @@ spec:
     namespace: paperless-ngx
   syncPolicy:
     automated:
-      enabled: false
+      enabled: true
       prune: false
-      selfHeal: false
+      selfHeal: true