ci: deploy every push to main through Carrel

.github/workflows/release.yml

@@ -73,3 +73,24 @@ jobs:
           labels: ${{ steps.meta.outputs.labels }}
           cache-from: type=gha,scope=embed
           cache-to: type=gha,mode=max,scope=embed
+
+  # Ships the new tag to Carrel: carrel-deployd on the machine runs
+  # deploy/carrel-start.sh and rolls back if web is unhealthy.
+  deploy:
+    needs: image
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      id-token: write
+    steps:
+      - uses: actions/checkout@v4
+      - run: |
+          mkdir release
+          cp deploy/carrel-start.sh release/
+          echo "sha-${GITHUB_SHA::7}" > release/VERSION
+          tar czf "$RUNNER_TEMP/gallery.tar.gz" -C release .
+      - uses: NYCU-WinLab/carrel-deploy@v1
+        with:
+          tenant: winfra
+          app: gallery
+          artifact: ${{ runner.temp }}/gallery.tar.gz

README.md

@@ -48,6 +48,8 @@ This starts Postgres, runs the migrations once, then serves the gallery from `gh
 > [!IMPORTANT]
 > Put a reverse proxy with TLS in front, at the host in `BETTER_AUTH_URL`, and let it pass request bodies of 100 MB. MCP clients only accept an HTTPS resource, and the Keycloak redirect URI must match it.
 
+gallery.winlab.tw deploys on every push to `main`. The release workflow pushes the image, then sends a Carrel app release (`winfra/gallery`). On the machine, carrel-deployd runs [deploy/carrel-start.sh](deploy/carrel-start.sh): it sets `GALLERY_VERSION` in `/opt/gallery/.env` and waits until web is healthy. A failed health check rolls back to the previous release. The app unit has a drop-in with `Type=notify`, `NotifyAccess=all` and `TimeoutStartSec=300`, so the check waits for the new container.
+
 ## Use
 
 1. In your Keycloak realm, create a confidential OpenID Connect client with the redirect URI `<BETTER_AUTH_URL>/api/auth/callback/keycloak`, and put its ID and secret in `.env`. Give its service account `view-users` so `/mcp` can check that a member still exists.

deploy/carrel-start.sh

@@ -0,0 +1,18 @@
+#!/bin/sh
+# carrel-deployd runs this from the release directory on every deploy and
+# rollback. It points the Compose stack in /opt/gallery at the image in
+# VERSION, reports ready once web is healthy, then follows the web logs.
+# The unit needs Type=notify and NotifyAccess=all (see README, Deploy).
+set -eu
+
+version=$(cat VERSION)
+cd /opt/gallery
+
+tmp=$(mktemp .env.XXXXXX)
+sed "s/^GALLERY_VERSION=.*/GALLERY_VERSION=$version/" .env >"$tmp"
+mv "$tmp" .env
+
+docker compose pull --quiet migrate web
+docker compose up -d --wait web
+systemd-notify --ready
+exec docker compose logs -f --since 1s web