Files
nixos/docs/superpowers/plans/2026-08-05-immich-nixos-module.md
T

300 lines
14 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Immich NixOS Module Implementation Plan
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
>
> **Note on nature:** Task 1 is repo work verifiable with `nix build` (no runtime tests exist for declarative config). Tasks 26 are a **manual migration runbook executed on jupiter by the operator** — they are destructive and cannot be run from the dev machine (mibook). Do not attempt to automate or execute Tasks 26 from an agent session; present them for the operator to run and confirm.
**Goal:** Replace jupiter's docker-compose Immich with the native `services.immich` NixOS module, preserving all data (albums, faces, shares, library).
**Architecture:** A standard `my.profiles.immich` module wraps `services.immich` (native Postgres+VectorChord over unix socket, Redis, server, machine-learning). Media stays at the default local `/var/lib/immich`. The existing docker Postgres dump is restored same-version (2.7.5 → 2.7.5, no schema/vector migration). GPU is exposed for VAAPI/QSV transcoding.
**Tech Stack:** NixOS (flake-parts), `services.immich` from nixpkgs 25.11, PostgreSQL, Intel QSV/VAAPI, docker (source only).
## Global Constraints
- Machine: **jupiter** only. Do not enable on mibook.
- Immich version: source docker == target nixpkgs == **2.7.5** (stable). No `package` override. Do NOT bump nixpkgs Immich during this work.
- Media location: default `/var/lib/immich` (local disk). Do not point at the NAS.
- Database: local PostgreSQL over **unix socket + peer auth** — no password, no sops secret.
- HW accel: **video transcoding only**. ML stays on CPU (`machine-learning.enable = true`, no OpenVINO).
- Access: LAN + VPN, `openFirewall = true`, port **2283**. No reverse proxy/TLS.
- Rebuild command: `sudo nixos-rebuild switch --flake '.#jupiter'`.
- Build-check command: `nix build '.#nixosConfigurations.jupiter.config.system.build.toplevel'`.
- Format Nix with `nixfmt-rfc-style` before committing.
- Do not delete docker DB or upload data until Task 6 sign-off.
---
## File Structure
- **Create** `modules/environments/immich/default.nix` — the `my.profiles.immich` module (single responsibility: declare Immich).
- **Modify** `modules/environments/default.nix` — add `./environments/immich` to the import list.
- **Modify** `machines/jupiter/environments.nix` — set `immich.enable = true`.
No other files change. The DB/media migration touches only runtime state on jupiter, not the repo.
---
### Task 1: Author the `immich` profile module
**Files:**
- Create: `modules/environments/immich/default.nix`
- Modify: `modules/environments/default.nix` (import list)
- Modify: `machines/jupiter/environments.nix` (`my.profiles.immich.enable`)
**Interfaces:**
- Produces: NixOS option `my.profiles.immich.enable` (bool). When true, configures `services.immich`, adds `immich` user to `video`/`render` groups, and appends an entry to `my.homepage.services`.
- Consumes: existing `my.homepage.services` aggregator; `config.networking.hostName`.
- [ ] **Step 1: Read a reference module to match repo style**
Read `modules/environments/jellyfin/default.nix` (same shape: `cfg`, `hostName`, `port`, `mkIf`, `my.homepage.services`). Match its formatting and header-comment convention.
- [ ] **Step 2: Create the module file**
Create `modules/environments/immich/default.nix`:
```nix
# Immich self-hosted photo & video server
{
config,
lib,
pkgs,
...
}:
let
cfg = config.my.profiles.immich;
hostName = config.networking.hostName;
port = 2283;
in
{
options.my.profiles.immich = with lib; {
enable = mkEnableOption "Immich photo server";
};
config = lib.mkIf cfg.enable {
services.immich = {
enable = true;
host = "0.0.0.0";
inherit port;
openFirewall = true;
mediaLocation = "/var/lib/immich";
machine-learning.enable = true;
accelerationDevices = [ "/dev/dri/renderD128" ];
settings.server.externalDomain = "http://${hostName}:${toString port}";
};
# The native module does not add GPU groups; required for VAAPI/QSV transcoding.
users.users.immich.extraGroups = [
"video"
"render"
];
my.homepage.services = [
{
group = "Media";
name = "Immich";
description = "Photo & video server";
href = "http://${hostName}:${toString port}";
icon = "immich.png";
}
];
};
}
```
- [ ] **Step 3: Register the module in the environments import list**
Open `modules/environments/default.nix` and add `./environments/immich` (or `./immich`, matching the exact relative style already used in that file — check how `jellyfin` is listed and mirror it).
- [ ] **Step 4: Enable it on jupiter**
In `machines/jupiter/environments.nix`, inside the `my.profiles = { ... }` block, add:
```nix
immich.enable = true;
```
- [ ] **Step 5: Format**
Run: `nixfmt-rfc-style modules/environments/immich/default.nix`
- [ ] **Step 6: Build-check (this is the "test")**
Run: `nix build '.#nixosConfigurations.jupiter.config.system.build.toplevel'`
Expected: builds successfully. If it fails on an unknown option (e.g. `accelerationDevices`, `settings.server.externalDomain`), reconcile against the module at `$(nix eval --raw '.#nixosConfigurations.jupiter.pkgs.path')/nixos/modules/services/web-apps/immich.nix` and fix.
- [ ] **Step 7: Confirm the option evaluates on**
Run: `nix eval '.#nixosConfigurations.jupiter.config.services.immich.enable'`
Expected: `true`
- [ ] **Step 8: Commit**
```bash
git add modules/environments/immich/default.nix modules/environments/default.nix machines/jupiter/environments.nix
git commit -m "feat(jupiter): add native Immich profile module"
```
---
### Task 2: Pre-flight & backup on jupiter (operator-run)
**Files:** none (runtime state on jupiter). Run all commands on jupiter.
**Interfaces:**
- Produces: `immich-db.sql` dump file and a known-good copy/snapshot of the docker upload folder; recorded `UPLOAD_LOCATION` path and DB container name.
- [ ] **Step 1: Record docker facts**
From the docker-compose dir on jupiter, note `UPLOAD_LOCATION`, the DB service/container name, and `POSTGRES_USER`/`POSTGRES_DB` from `.env`/compose. Confirm server version is **2.7.5** (web UI footer or `docker exec <server> immich --version`). If it is not 2.7.5, STOP — this plan assumes a same-version restore.
- [ ] **Step 2: Stop the docker stack (DB may stay up for the dump)**
Run: `docker compose stop immich-server immich-machine-learning` (leave the DB container running).
- [ ] **Step 3: Dump the database**
Run: `docker exec -t <db-container> pg_dumpall --clean --if-exists --username=<POSTGRES_USER> > ~/immich-db.sql`
Expected: a non-trivial `immich-db.sql` (check it is not near-empty: `wc -l ~/immich-db.sql`).
- [ ] **Step 4: Stop the DB and record the media size**
Run: `docker compose down` then `du -sh <UPLOAD_LOCATION>` and note the size. Do NOT copy yet. Do NOT delete anything.
---
### Task 3: First switch — let the module create empty state (operator-run)
**Files:** none at runtime (repo change already committed in Task 1). Run on jupiter after pulling the committed branch.
**Interfaces:**
- Consumes: `immich-db.sql`, `UPLOAD_LOCATION` from Task 2.
- Produces: an `immich` system user, an empty `immich` Postgres DB + role, and `/var/lib/immich` created with correct ownership, with services then stopped.
- [ ] **Step 1: Deploy the config**
On jupiter, check out the branch containing Task 1's commit and run:
`sudo nixos-rebuild switch --flake '.#jupiter'`
Expected: `immich-server`, `immich-machine-learning`, postgres, and redis units come up; UI reachable at `http://jupiter:2283` showing a fresh/empty instance.
- [ ] **Step 2: Stop immich so data can be swapped underneath**
Run: `sudo systemctl stop immich-server immich-machine-learning`
Expected: both inactive. PostgreSQL and Redis stay running.
- [ ] **Step 3: Verify the DB and user exist**
Run: `sudo -u postgres psql -c '\l' | grep immich` and `sudo -u postgres psql -c '\du' | grep immich`
Expected: an `immich` database and `immich` role are present.
---
### Task 4: Restore database and media (operator-run, destructive)
**Files:** none in repo. Run on jupiter. This overwrites the freshly-created empty DB.
**Interfaces:**
- Consumes: `immich-db.sql`, `<UPLOAD_LOCATION>`, the running NixOS PostgreSQL.
- Produces: the migrated DB and populated `/var/lib/immich`.
- [ ] **Step 1: Restore the dump into the NixOS Postgres**
`pg_dumpall` output includes role/DB creation. Load it as the `postgres` superuser over the unix socket:
Run: `sudo -u postgres psql -f ~/immich-db.sql`
Expected: completes without fatal errors. Harmless "role already exists"/"database already exists" notices are OK because of `--clean --if-exists`. If the immich DB ends up owned by the wrong role, reassign: `sudo -u postgres psql -c 'ALTER DATABASE immich OWNER TO immich;'`.
- [ ] **Step 2: Sanity-check the restored data**
Run: `sudo -u postgres psql -d immich -c 'SELECT count(*) FROM assets;'`
Expected: a count matching your library size (non-zero). If the table name differs by version, list tables with `\dt` and check an obviously-populated one.
- [ ] **Step 3: Move the media into the default location**
Immich's upload folder holds subdirs `library/ upload/ thumbs/ encoded-video/ profile/ backups/`. Move (not copy, if same filesystem) the contents of `<UPLOAD_LOCATION>` into `/var/lib/immich`:
Run: `sudo rsync -aHAX --info=progress2 <UPLOAD_LOCATION>/ /var/lib/immich/`
(Use `rsync` — safe if partially interrupted. Keep the source until Task 6 sign-off.)
- [ ] **Step 4: Fix ownership**
Run: `sudo chown -R immich:immich /var/lib/immich`
Expected: everything under `/var/lib/immich` owned by `immich`.
---
### Task 5: Start and verify (operator-run)
**Files:** none. Run on jupiter.
**Interfaces:**
- Consumes: migrated DB + media from Task 4.
- Produces: a running, verified native Immich.
- [ ] **Step 1: Start the server and watch logs**
Run: `sudo systemctl start immich-server && journalctl -u immich-server -f`
Expected: it connects to the DB, runs same-version startup checks (no destructive migration since 2.7.5==2.7.5), and reports listening on 2283. Leave the follow running through the next step.
- [ ] **Step 2: Start machine-learning**
Run: `sudo systemctl start immich-machine-learning`
Expected: active, no crash loop in `journalctl -u immich-machine-learning`.
- [ ] **Step 3: Functional spot-check in the web UI**
At `http://jupiter:2283`: log in with an existing account; confirm the timeline loads; open an **album**; open the **People/faces** view; open a **shared link**; open one photo so a **thumbnail and its full original both load** (this proves DB↔file paths align after the media move).
Expected: all present, images render.
- [ ] **Step 4: Confirm homepage dashboard tile**
Open the homepage dashboard; confirm the Immich tile appears under "Media" and links to `http://jupiter:2283`.
- [ ] **Step 5: Enable and verify hardware transcoding**
In Immich **Administration → Settings → Video Transcoding**, set hardware acceleration to **Quick Sync** (QSV) (or VAAPI). Trigger a transcode (upload/play a video that needs transcoding, or run the transcoding job). Then:
Run: `journalctl -u immich-server | grep -iE 'qsv|vaapi|hwaccel|transcode'`
Expected: log shows the hardware path in use, not a CPU-fallback error. Confirm `/dev/dri/renderD128` is accessible to the service (the `video`/`render` groups + `accelerationDevices` from Task 1 handle this).
---
### Task 6: Sign-off and cleanup (operator-run)
**Files:** none in repo. Merge the branch; then, only after a confidence window, remove docker.
**Interfaces:**
- Consumes: a verified running instance (Task 5).
- [ ] **Step 1: Merge the feature branch**
Open a PR from `feat/immich-nixos-module` into `main` and merge it (repo convention: PRs via the Gitea remote).
- [ ] **Step 2: Confidence window**
Use Immich normally for a few days. Keep the docker `<UPLOAD_LOCATION>` source copy and `~/immich-db.sql` untouched as the rollback path.
- [ ] **Step 3: Rollback (only if needed, before cleanup)**
If something is wrong: `sudo systemctl stop immich-server immich-machine-learning`, set `immich.enable = false` (or check out the pre-migration commit), `sudo nixos-rebuild switch --flake '.#jupiter'`, then `docker compose up -d` in the old stack. Original docker DB + upload folder are intact until Step 4.
- [ ] **Step 4: Cleanup (after sign-off)**
Remove the docker Immich stack (`docker compose down --rmi all --volumes` in the old dir if the DB volume is dedicated — verify first), delete the now-duplicated `<UPLOAD_LOCATION>` source, and remove `~/immich-db.sql`. Optionally disable the `docker` profile on jupiter if Immich was its only consumer (check other services first — jupiter's `docker.enable` may still be needed).
---
## Self-Review
**Spec coverage:**
- Native `services.immich` → Task 1. ✓
- Version target 2.7.5==stable, no override → Global Constraints + Task 2 Step 1. ✓
- Media at default `/var/lib/immich` → Task 1 + Task 4 Step 3. ✓
- DB migrate keep-everything → Tasks 24. ✓
- HW transcoding only → Task 1 (`accelerationDevices`, groups) + Task 5 Step 5. ✓
- LAN+VPN, port 2283, openFirewall, homepage tile → Task 1 + Task 5 Steps 34. ✓
- Rollback path → Task 6 Step 3. ✓
- Deferred (OpenVINO/NAS/proxy) → correctly absent. ✓
**Placeholder scan:** No TBD/TODO; every command is concrete. Placeholders like `<db-container>`, `<UPLOAD_LOCATION>`, `<POSTGRES_USER>` are runtime values the operator reads in Task 2 Step 1 — intentional, not gaps.
**Type consistency:** Option name `my.profiles.immich.enable` and path `/var/lib/immich` used consistently across all tasks. Media subfolder list matches between Task 4 Step 3 and the spec.