Skip to content

Pick the reverse proxy per machine: Caddy, nginx, Apache or Traefik - #20

Merged
liamdevss merged 1 commit into
mainfrom
proxy-choice
Nov 16, 2026
Merged

liamdevss merged 1 commit into
mainfrom
proxy-choice

Conversation

@liamdevss

Copy link
Copy Markdown
Contributor

For serveroshq/panel#92. A deploy can now say which reverse proxy should serve the machine's domains. If it isn't the one in use, the daemon moves every site ServerOS serves on the machine over to it first, and keeps the choice in daemon.toml.

What each proxy does

  • Caddy: unchanged. Automatic HTTPS, wildcards through on-demand TLS.
  • nginx: sites in /etc/nginx/serveros.d, included from the http block. HTTPS through certbot (webroot challenge in /var/lib/serveros/acme), with an HTTP to HTTPS redirect once the certificate is there and a renewal hook that reloads nginx. Fixes the old renderer, which used $connection_upgrade without defining it, so nginx -t failed on a stock config.
  • Apache: sites in /etc/apache2/serveros.d through IncludeOptional, the proxy modules switched on, HTTPS through certbot like nginx.
  • Traefik: installed from the latest v3 GitHub release (checked against its checksums), run as its own user with only CAP_NET_BIND_SERVICE. Sites are TOML files for its file provider and get certificates from a serveros ACME resolver. A Traefik someone set up by hand is left alone: the deploy says what to add to its config.

Wildcard subdomains stay plain HTTP on nginx, Apache and Traefik (the deploy says so): a certificate per subdomain needs Caddy's on-demand TLS. Behind a tunnel, every proxy serves plain HTTP.

Switching
Stops the current proxy (only one can have ports 80 and 443), installs the new one if needed, writes each site for it and starts it. If anything fails on the way, the new site files are removed and the old proxy is started again. Site files now carry a # ServerOS site: port=… domains=… line so any proxy can read them back; Caddy and nginx files written before it still read back.

Certificates that aren't ready yet
certbot doesn't keep retrying the way Caddy does. A site whose domain doesn't point here yet stays on plain HTTP, and a worker tries again every 15 minutes.

Wire changes (all optional, older panels and daemons ignore them)

  • DeploySpec.proxy: caddy, nginx, apache or traefik.
  • Hello.proxy: the proxy in use.
  • The deploy result has proxy.

Checked

  • cargo fmt, cargo clippy --workspace --all-targets -D warnings, cargo test --workspace.
  • The rendered configs loaded into real servers in containers: nginx and Apache from Debian bookworm (nginx -t, apache2ctl configtest, then the challenge path, the redirect, HTTPS proxying, wildcard and tunnel sites), and traefik:v3 (redirect, wildcard and tunnel routing, no config errors).
  • Not run end to end: a switch on a real machine with systemd.

…with HTTPS for each, and switching moves the machine's existing sites over
@liamdevss
liamdevss requested a review from a team as a code owner November 16, 2026 18:53
liamdevss added a commit to serveroshq/panel that referenced this pull request Nov 16, 2026
Closes #92. Goes with serveroshq/daemon#20.

When a deploy has a domain, the deploy dialog now has a **Reverse
proxy** dropdown: Caddy, nginx, Apache or Traefik. It starts on whatever
the machine uses. Picking another one shows that deploying moves every
site on the machine over and stops the old proxy, since only one of them
can have ports 80 and 443.

**Panel**
- `machines.proxy`: what serves the machine's domains, from the daemon's
hello and from each finished deploy.
- `/machines/{machine}/deploy/taken` also returns `proxy` and
`canChooseProxy`, which is how the dialog learns both when it opens.
- A deploy sends `proxy` only when it differs from the machine's current
one and there's a domain. Machines that never reported a proxy count as
Caddy.
- Gated on `daemon.proxy_choice_since` (0.1.1-pre.40, the next daemon
release). Older daemons don't get the dropdown, and asking for a proxy
through the API or MCP is turned down with "too old to switch reverse
proxy".
- `deploy_from_git` and `deploy_template` (MCP) take `proxy` too.
- Copy follows the proxy: the wildcard hint says subdomains are plain
HTTP on nginx, Apache and Traefik, and the done step names the proxy
that gets the certificate. Translations added for es, de and hi.
- A proxy that won't start is explained by name, with the right
`systemctl` command.

**Checked**
- New `ReverseProxyChoiceTest` (10 tests), and `ServiceTemplateTest`
updated for the two new fields on the taken endpoint.
- Full `tests/` suite: one failure, `SyncDaemonReleasesTest`, which also
fails without this change in my local setup (local `.env` release URL).
- Pint, PHPStan on the changed files, `vp check`, `vue-tsc`.
- Not checked: the dropdown rendered in a browser.

Co-authored-by: liamdevss <213730082+liamdevss@users.noreply.github.com>
@liamdevss
liamdevss merged commit b81d603 into main Nov 16, 2026
12 checks passed
@liamdevss
liamdevss deleted the proxy-choice branch November 16, 2026 19:07
liamdevss added a commit that referenced this pull request Nov 16, 2026
Follow-up to #20, found by testing a switch on a real systemd machine.

When the new proxy started and then fell over a moment later (in the
test: something else had port 443, so Traefik exited right after
`systemctl enable --now` returned), the switch counted it as done. The
old proxy was already stopped and was then disabled, so every site went
offline.

Now, after writing the sites, the switch waits 4 seconds and checks that
the new proxy's unit is still active and that port 80 accepts a
connection. If not, it fails with the unit's journal, removes the new
site files, stops **and disables** the new proxy (so it doesn't fight
for the ports on the next boot), and starts the old one again.

**Checked on Ubuntu 24.04 with systemd (a privileged container):**
- Every switch with two sites: Caddy → nginx → Apache → Traefik → Caddy
→ nginx. Each time the new proxy holds port 80, both sites answer and
the old proxy is inactive and disabled.
- With port 443 held by another program, nginx → Traefik is refused
("Traefik started but stopped again"). nginx is active and enabled
again, both sites answer, and Traefik is failed and disabled.
- `cargo fmt`, `cargo clippy --workspace --all-targets -D warnings`, and
the daemon-deploy tests.

Co-authored-by: liamdevss <213730082+liamdevss@users.noreply.github.com>
liamdevss added a commit that referenced this pull request Nov 16, 2026
…#23)

Fixes serveroshq/panel#97.

**The bug.** certbot's HTTP challenge files went in
`/var/lib/serveros/acme`, inside the daemon's state folder, which is
`0700` root. nginx and Apache workers run as `www-data`, so Let's
Encrypt's request for the challenge file got a 403 and every certificate
request failed. Sites stayed on plain HTTP, retrying every 15 minutes
forever. (My container test in #20 never got this far: Let's Encrypt
refuses `.test` names before it asks for the file.)

**The fix**
- The challenge folder is now `/var/lib/serveros-acme`, created `0755`.
- At start-up (30 s in), the daemon rewrites any site file it would now
write differently and reloads the proxy, putting the old files back if
it won't take them. Machines on v0.1.1-pre.40 fix themselves after
updating, and the next certificate retry succeeds.

**HTTPS on or off, per site**
- `DeploySpec.https` (optional): false serves the domains over plain
HTTP on every proxy, with no certificate and no redirect. That's for a
CDN or proxy in front that does HTTPS, or a site that shouldn't have it.
- The choice is kept in the site file (`https=off` on the `# ServerOS
site:` line), so rollbacks, clones and switching proxies keep it. A
deploy that doesn't say keeps what the site had, and HTTPS for a new
one.

**Checked on Ubuntu 24.04 with systemd**
- With `/var/lib/serveros` at `700`, the old layout gives a 403 for the
challenge file (workers run as www-data). The new one serves it.
- A site file written by pre.40 gets its challenge folder rewritten by
`refresh_sites`, and the challenge is served afterwards.
- A site with HTTPS off is served over HTTP with `https=off` in its
file, and certbot isn't run for it.
- `cargo fmt`, `cargo clippy --workspace --all-targets -D warnings` and
the tests. New tests cover the folder location, HTTPS off on all four
proxies (nothing on 443, no redirect, no resolver, no on-demand TLS),
reading it back, and a certificate on disk not turning an HTTPS-off site
back on.

Co-authored-by: liamdevss <213730082+liamdevss@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant