* A WROLPi serves its library over rsync, and makes a database snapshot on request
The first part of replacing the sync: a secondary will copy its primary's whole media directory with rsync and
replace its database with the primary's, instead of downloading the files of FileGroups over HTTP (which never
copies an ignored directory, such as the map's).
- An rsync daemon serves the media directory read-only on port 8873, on every WROLPi. It never serves the TLS keys,
the cookies, or the database in use. On a Raspberry Pi and Debian it is the wrolpi-rsyncd unit; in Docker the
Controller runs it in a helper container made from the api container (its image, user and volumes), so the media
directory is still configured in one place. All three read controller/rsyncd.conf.
- POST /api/sync/snapshot asks for a snapshot of the database, which the daemon serves from config/.sync. An
unchanged database is not written again.
- /api/sync/info also says the database's migration and whether the media directory is mounted.
Nothing uses these yet; the existing sync is unchanged.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
* The rsync daemon follows no links, and no snapshot is made in WROL Mode
- A client that asked the daemon to follow symbolic links could read what a link in the library points at. A
relative link to config/ssl was read this way although that directory is left out of the module. The daemon now
refuses the three options that follow links; a link is sent as a link.
- A database snapshot is a new file on the primary's drive, so WROL Mode refuses it, as it does a backup. A snapshot
made earlier is still served.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
* No rsync helper from an image without rsync, and the helper follows the api container
Review found:
- A helper made from an api image built before rsync was added exits at once, and Docker restarted it forever while
the Controller reported it created. The image is now looked at first; no helper is made from one without rsync,
an older helper is left serving, and the reason is logged once.
- The helper was only checked when the Controller started, so a recreated or rebuilt api container left it stale.
It is now checked every minute.
- A snapshot cut off by its process being killed stayed marked as running and held back every later request.
- The snapshot's half-written temporary file was served by the daemon.
- The comment on the daemon's refused options claimed more than they give: a link that is part of the requested
path is still followed.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
Drop .pgpass handling, Postgres ownership repairs, and the docker db
container viewability rules. Update comments and pi-gen map docs for
SQLite/PMTiles. Help submodule documents the new database layout.
- Replace nginx with Caddy as the web server for both Docker and native
(RPi/Debian) deployments. Caddy handles HTTPS termination, reverse
proxying to all services (API, React app, Controller, Kiwix, Help, Map),
and serves the media/download file browsers.
- Generate a Root CA and signed leaf certificates for HTTPS. The Root CA
uses Name Constraints to restrict signing to local addresses only
(.local, localhost, and RFC 1918 private IP ranges), mitigating MITM
risk if the CA key is compromised.
- Add an HTTP landing page that guides users through one-time CA
certificate installation on iOS, Android, macOS, Windows, Linux, and
Firefox, with auto-redirect when the cert is already trusted.
- Add trailing-slash redirects for /controller, /docs, /media, and
/download so bare paths don't fall through to the React catch-all.
- Add a Root CA download link on the Settings page under Configs.
- Update help.sh diagnostics for IPv6 listener detection and HTTP/2
status codes.
- Remove all nginx config files and the https_proxy Docker container.
- Bump wrolpi-help submodule with certificate documentation.