Skip to content

Sensitive tier: Paperless-ngx #133

Closed

Description

@Gerrrt

Sub-issue of #102.

What and why

Paperless-ngx GPL-3.0 Python/Docker — awesome-selfhosted README.md:755.

Scan, OCR, index and archive paper documents. The maintained fork of Paperless; the original and Paperless-ng are both dormant, so this is the only one worth deploying.

It earns its place on the sensitive tier by content rather than by size: a document archive is tax returns, insurance, medical records, passports and payslips. docs/security.md already reasons this way about what not to publish; the same logic puts this behind the same boundary as Vaultwarden (#131).

What it needs

  • OCR is bursty CPU. Tesseract on a backlog scan will use everything it is given for minutes at a time. That is fine on an N100 and is one of the reasons the mini PC is being bought, but it makes a memory and CPU limit worth setting from day one rather than after the first OOM (Set memory limits on the six services to bound the blast radius of a leak #114, No resource limits on any container, and no size bound on retention #71).
  • Postgres, Redis, and a consume directory.
  • Secret key and database credentials in secrets/*.sops.yaml.
  • A document-ingest path from a workstation on Hicks. The existing 50→99 rule already covers reaching it; no new rule.
  • Storage sizing, and a backup that covers the archive and the database. An OCR index can be rebuilt; the originals cannot.

A note on scope

Paperless is a document archive, not a document editor, and it deliberately stops there. Stirling-PDF is filed separately for the merge/split/convert operations that otherwise send household paperwork to a random website — the two pair well, and neither wants to be the other.

Which VLAN this touches

🔴 99 — Winterfell (management)

Security review

  • Does not weaken the default-deny posture between VLANs — reached from Hicks under the existing 50→99; no new rule
  • Any new credential goes in secrets/*.sops.yaml
  • Any new device is added to docs/network.md with an OUI-truncated MAC — n/a, unless a network scanner is added later, which would put a device with hard-coded credentials on a segment worth thinking about first

Depends on

#129 (Caddy) and the mini PC in #102. Related: #122.


Corrected 2026-09-19

Authored as three services in #430: paperless, paperless-db, paperless-broker — Valkey, not Redis — with four volumes, all sentinelled, including the version-carrying ./18/docker/PG_VERSION. The tier's first CPU ceiling. "Depends on #102" → #404.

Acceptance beyond starting: the 4 cores and 3 GiB for OCR are, per stacks/sensitive/README.md, "a statement about the ProDesk made on a different machine". Measure them on trinity under a real scan before calling this done.

Activity

  1. added 2 commits that reference this issue on Sep 10, 2026
  2. added this to the Wave 1 — trinity milestone on Sep 11, 2026
  3. Gerrrt commented on Sep 28, 2026

    @Gerrrt
    OwnerAuthor

    Accepted on trinity — recorded in #700 (ee17801).

    Under a synthetic backlog (five one-page scans and one of 50 pages, 300 dpi, image-only): memory.peak 1716 MiB of 3072m with no OOM kill, 3.0 cores at the busiest minute and 0.2 s throttled in total, 3 min 54 s for the 50 pages, and Vaultwarden through Caddy never slower than 19 ms. Both limits stand; the one-month container_memory_rss re-derivation still applies. The test documents were deleted afterwards.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    featureA new capability or a changesensitivestacks/sensitive on trinity (VLAN 99)

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions