selfhosted-apps-docker/caddy_v2

Caddy v2 Reverse Proxy

guide-by-example

1. Purpose & Overview
2. Caddy as a reverse proxy in docker
3. Caddy more info and various configurations
4. Caddy DNS challenge

Purpose & Overview

Reverse proxy is needed if one wants access to services based on the hostname.
For example nextcloud.example.com points traffic to nextcloud docker container, while jellyfin.example.com points to the media server on the network.

• Official site
• Official documentation
• Forum
• Github

Caddy is a pretty damn good web server with automatic HTTPS. Written in Go.
Web servers are build to deal with http traffic, so they are an obvious choice for the function of reverse proxy.
In this setup Caddy is used mostly as a TLS termination proxy. Https encrypted tunel ends with it, so that the traffic can be analyzed and send to a correct webserver based on the settings in Caddyfile.

Caddy with its build-in https and and simple config approach allows even most trivial configs to just work:

nextcloud.{$MY_DOMAIN} {
  reverse_proxy nextcloud-web:80
}

jellyfin.{$MY_DOMAIN} {
  reverse_proxy 192.168.1.20:80
}

And just works means fully works. No additional configuration needed for https redirect, or special services if target is not a container, or need to deal with load balancer, or need to add boilerplate headers for x-forward, or other extra work.
It has great out of the box defaults, fitting majority of uses and only some special casess with extra functionality need extra work.

Caddy as a reverse proxy in docker

Caddy will be running as a docker container and will route traffic to other containers, or machines on the network.

- Requirements

• have some basic linux knowledge, create folders, create files, edit files, run scripts,...
• have a docker host and some vague docker knowledge
• have port 80 and 443 forwarded on the router/firewall to the docker host
• have a domain, example.com, you can buy one for 2€ annually on namecheap
• have correctly set type-A DNS records pointing at your public IP address, switching to Cloudflare for DNS managment is recommended

- Files and directory structure

/home/
└── ~/
    └── docker/
        └── caddy/
            ├── config/
            ├── data/
            ├── .env
            ├── Caddyfile
            └── docker-compose.yml

• config/ - a directory containing configs that Caddy generates, most notably autosave.json which is a backup of the last loaded config
• data/ - a directory storing TLS certificates
• .env - a file containing environment variables for docker compose
• Caddyfile - the Caddy configuration file
• docker-compose.yml - a docker compose file, telling docker how to run containers

You only need to provide the three files.
The directories are created by docker compose on the first run, the content of these is visible only as root of the docker host.

- Create a new docker network

docker network create caddy_net

All the containers and Caddy must be on the same network.

- Create .env file

You want to change example.com to your domain.

.env

MY_DOMAIN=example.com
DOCKER_MY_NETWORK=caddy_net

Domain names, api keys, email settings, ip addresses, database credentials, ... whatever is specific for one deployment and different for another, all of that ideally goes in to the .env file.

If .env file is present in the directory with the compose file, it is automatically loaded and these variables will be available for docker-compose when building the container with docker-compose up. This allows compose files to be moved from system to system more freely and changes are done to the .env file.

Often variable should be available also inside the running container. For that it must be declared in the environment section of the compose file, as can be seen next in Caddie's docker-compose.yml

extra info:
docker-compose config shows how compose will look with the variables filled in.

- Create docker-compose.yml

docker-compose.yml

services:

  caddy:
    image: caddy
    container_name: caddy
    hostname: caddy
    restart: unless-stopped
    ports:
      - "80:80"
      - "443:443"
    environment:
      - MY_DOMAIN
    volumes:
      - ./Caddyfile:/etc/caddy/Caddyfile
      - ./data:/data
      - ./config:/config

networks:
  default:
    name: $DOCKER_MY_NETWORK
    external: true

• port 80 and 443 are pusblished for http and https
• MY_DOMAIN variable is passed in to the container so that it can be used in Caddyfile
• the Caddyfile is bind-mounted from the docker host
• directories data and config are bind mounted so that their content persists
• the same network is joined as for all other containers

- Create Caddyfile

Caddyfile

{
    # acme_ca https://acme-staging-v02.api.letsencrypt.org/directory
}

a.{$MY_DOMAIN} {
    reverse_proxy whoami:80
}

b.{$MY_DOMAIN} {
    reverse_proxy nginx:80
}

a and b are the subdomains a.example.com and b.example.com, can be named whatever. For them to work they must have type-A DNS record pointing at your public ip set on Cloudflare, or wherever the domains DNS is managed.

The value of {$MY_DOMAIN} is provided by the compose and the .env file.
The subdomains point at docker containers by their hostname and exposed port. So every docker container you spin should have hostname definied.
Commented out is the staging url for let's encrypt.

- Setup some docker containers

Something light and easy to setup to route to.
Assuming for this testing these compose files are in the same directory with Caddy, so they make use of the same .env file and so be on the same network.

Note the lack of published/mapped ports in the compose, as they will be accessed only through Caddy, which has it's ports published.
And since the containers and Caddy are all on the same bridge docker network, they can access each other on any port.
Exposed ports are just documentation, don't confuse expose and publish.

extra info:
To know which ports containers have exposed - docker ps, or docker port <container-name>, or use ctop.

whoami-compose.yml

version: "3.7"
services:

  whoami:
    image: "containous/whoami"
    container_name: "whoami"
    hostname: "whoami"

networks:
  default:
    name: $DOCKER_MY_NETWORK
    external: true

nginx-compose.yml

version: "3.7"
services:

  nginx:
    image: nginx:latest
    container_name: nginx
    hostname: nginx

networks:
  default:
    name: $DOCKER_MY_NETWORK
    external: true

- editing hosts file

You are on your local network and you are likely running the docker host inside the same network.
If that's the case then shit will not work without editing the hosts file.
Reason being that when you write that a.example.com in to your browser, you are asking google's DNS for a.example.com IP address. It will give you your own public IP, and most routers/firewalls wont allow this loopback, where your requests should go out and then right back.

So just edit hosts as root/administrator, adding whatever is the local IP of the docker host and the hostname:

192.168.1.222     a.{$MY_DOMAIN}
192.168.1.222     b.{$MY_DOMAIN}

If it is just quick testing one can use Opera browser and enable the build in VPN.

One can also run a dns/dhcp server on the network, to solve this for all devices.
Here's a guide-by-example for dnsmasq.

- Run it all

Caddy

• docker-compose up -d

Services

• docker-compose -f whoami-compose.yml up -d
• docker-compose -f nginx-compose.yml up -d

Give it time to get certificates, checking docker logs caddy as it goes, then visit the urls. It should lead to the services with https working.

If something is fucky use docker logs caddy to see what is happening.
Restarting the container docker container restart caddy can help. Or investigate inside docker exec -it caddy /bin/sh. For example trying to ping hosts that are suppose to be reachable, ping nginx should work.

There's also other possible issues, like bad port forwarding towards docker host.

extra info:
docker exec -w /etc/caddy caddy caddy reload reloads config if you made changes and want them to take effect.

extra info2:
caddy can complain about formatting of the Caddyfile
this executed on the host will let caddy overwrite the Caddyfile with correct formatting docker exec -w /etc/caddy caddy caddy fmt -overwrite

Caddy more info and various configurations

Caddyfile structure:

Worth having a look at the official documentation, especially these short pages

• concept
• conventions
• reverse_proxy

Maybe checking out mozzila's - overview of HTTP would also not hurt, it is very well written.

Routing traffic to other machines on the LAN

If not targeting a docker container but a dedicated machine on the network.
Nothing really changes, if you can ping the machine from Caddy container by its hostname or its IP, it will work.

blue.{$MY_DOMAIN} {
  reverse_proxy server-blue:80
}

violet.{$MY_DOMAIN} {
  reverse_proxy 192.168.1.100:80
}

Reverse proxy without domain and https

You can always just use localhost, which will translates in to docker hosts IP address.

localhost:55414 {
  reverse_proxy urbackup:55414
}

:9090 {
  reverse_proxy prometheus:9090
}

Prometheus entry uses short-hand notation.
TLS is automatically disabled in localhost use.

But for this to work Caddy's compose file needs to have those ports published too.

docker-compose.yml

services:

  caddy:
    image: caddy
    container_name: caddy
    hostname: caddy
    restart: unless-stopped
    ports:
      - "80:80"
      - "443:443"
      - "55414:55414"
      - "9090:9090"
    environment:
      - MY_DOMAIN
    volumes:
      - ./Caddyfile:/etc/caddy/Caddyfile:ro
      - ./data:/data
      - ./config:/config

networks:
  default:
    name: $DOCKER_MY_NETWORK
    external: true

With this setup, and assuming docker host at: 192.168.1.222, writing 192.168.1.222:55414 in to browser will go to to urbackup, and 192.168.1.222:9090 gets to prometheus.

Named matchers and IP filtering

Caddy has matchers which allow you to define how to deal with incoming requests.
reverse_proxy server-blue:80 is a matcher that matches all requests and sends them somewhere.
But if more control is desired, path matchers and named matchers come to play.

What if you want to block all traffic coming from the outside world, but local network be allowed through?
Well, the remote_ip matcher comes to play, which enables you to filter requests by their IP.

Named matchers are defined by @ and can be named whatever you like.

{
    # acme_ca https://acme-staging-v02.api.letsencrypt.org/directory
}

a.{$MY_DOMAIN} {
    reverse_proxy whoami:80
}

b.{$MY_DOMAIN} {
    reverse_proxy nginx:80

    @fuck_off_world {
        not remote_ip 192.168.1.0/24
    }
    respond @fuck_off_world 403
}

@fuck_off_world matches all IPs except the local network IP range.
Requests matching that rule get the response 403 - forbidden.

Snippets

What if you need to have the same matcher in several site-blocks and would prefer for config to look cleaner?

Here comes the snippets.
Snippets are defined under the global options block, using parentheses, named whatever you like.
They then can be used inside any site-block with simple import <snippet name>

Now would be a good time to look again at that concept picture above.

Here is above example of IP filtering named matcher done using a snippet.

{
    # acme_ca https://acme-staging-v02.api.letsencrypt.org/directory
}

(LAN_only) {
    @fuck_off_world {
        not remote_ip 192.168.1.0/24
    }
    respond @fuck_off_world 403
}

a.{$MY_DOMAIN} {
    reverse_proxy whoami:80
}

b.{$MY_DOMAIN} {
    reverse_proxy nginx:80
    import LAN_only
}

Backend communication

Some containers might be set to communicate only through https 443 port. But since they are behind proxy, their certificates wont be singed, wont be trusted.

Caddies sub-directive transport sets how to communicate with the backend.
Setting the upstream's scheme to https:// or declaring the tls transport subdirective makes it use https. Setting tls_insecure_skip_verify makes Caddy ignore errors due to untrusted certificates coming from the backend.

whatever.{$MY_DOMAIN} {
    reverse_proxy https://server-blue:443 {
        transport http {
            tls
            tls_insecure_skip_verify
        }
    }
}

HSTS and redirects

Here is an example of a redirect when wanting the common case of switching anyone that comes to a www subdomain to the naked domain.

www.{$MY_DOMAIN} {
    redir https://{$MY_DOMAIN}{uri}
}

Another example is running NextCloud behind proxy, which likely shows few warning on its status page. It requires some redirects for service discovery to work and would like if HSTS 2 would be set.
Like so:

nextcloud.{$MY_DOMAIN} {
    reverse_proxy nextcloud:80
    header Strict-Transport-Security max-age=31536000;
    redir /.well-known/carddav /remote.php/carddav 301
    redir /.well-known/caldav /remote.php/caldav 301
}

Headers and gzip

This example is with bitwarden_rs password manager, which comes with its reverse proxy recommendations.

encode gzip enables compression.
This lowers the bandwith use and speeds up loading of the sites. It is often set on the webserver running inside the docker container, but if not it can be enabled on caddy. You can check if your stuff has it enabled by using one of many online tools

By default, Caddy passes through Host header and adds X-Forwarded-For for the client IP. This means that 90% of the time a simple config is all that is needed but sometimes some extra headers might be desired.

Here we see bitwarden make use of some extra headers.
We can also see its use of websocket protocol for notifications at port 3012.

bitwarden.{$MY_DOMAIN} {
    encode gzip

    header {
        # Enable cross-site filter (XSS) and tell browser to block detected attacks
        X-XSS-Protection "1; mode=block"
        # Disallow the site to be rendered within a frame (clickjacking protection)
        X-Frame-Options "DENY"
        # Prevent search engines from indexing (optional)
        X-Robots-Tag "none"
        # Server name removing
        -Server
    }

    # Notifications redirected to the websockets server
    reverse_proxy /notifications/hub bitwarden:3012

    # Proxy the Root directory to Rocket
    reverse_proxy bitwarden:80
}

Basic authentication

Official documentation.
Directive basicauth can be used when one needs to add a username/password check before accessing a service.

Password is bcrypt hashed and then base64 encoded.
You can use the caddy hash-password command to hash passwords for use in the config.

Config bellow has login/password : bastard/bastard

Caddyfile

b.{$MY_DOMAIN} {
    reverse_proxy whoami:80
    basicauth {
        bastard JDJhJDA0JDVkeTFJa1VjS3pHU3VHQ2ZSZ0pGMU9FeWdNcUd0Wk9RdWdzSzdXUXNhWFFLWW5pYkxXVEU2
    }
}

Logging

Official documentation.
If access logs for specific site are desired

bookstack.{$MY_DOMAIN} {
    log {
        output file /data/logs/bookstack_access.log {
            roll_size 20mb
            roll_keep 5
        }
    }
    reverse_proxy bookstack:80
}

Caddy DNS challenge

This setup only works for Cloudflare.

DNS challenge authenticates ownership of the domain by requesting that the owner puts a specific TXT record in to the domains DNS zone.
Benefit of using DNS challenge is that there is no need for your server to be reachable by the letsencrypt servers. Cant open ports or want to exclude entire world except your own country from being able to reach your server? DNS challange is what you want to use for https then.
It also allows for issuance of wildcard certificates.
The drawback is a potential security issue, since you are creating a token that allows full control over your domain's DNS. You store this token somewhere, you are giving it to some application from dockerhub...

note: caddy uses a new libdns golang library with cloudflare package

- Create API token on Cloudflare

On Cloudflare create a new API Token with two permsisions, pic of it here

• zone/zone/read
  
• zone/dns/edit
  

Include all zones needs to be set.

- Edit .env file

Add CLOUDFLARE_API_TOKEN variable with the value of the newly created token.

.env

MY_DOMAIN=example.com
DOCKER_MY_NETWORK=caddy_net

CLOUDFLARE_API_TOKEN=<cloudflare api token goes here>

- Create Dockerfile

To add support, Caddy needs to be compiled with Cloudflare DNS plugin.
This is done by using your own Dockerfile, using the builder image.

Create a directory dockerfile-caddy in the caddy directory.
Inside create a file named Dockerfile.

Dockerfile

FROM caddy:2.6.2-builder AS builder

RUN xcaddy build \
    --with github.com/caddy-dns/cloudflare

FROM caddy:2.6.2

COPY --from=builder /usr/bin/caddy /usr/bin/caddy

- Edit docker-compose.yml

image replaced with build option pointing at the Dockerfile location
and CLOUDFLARE_API_TOKEN variable added.

docker-compose.yml

services:

  caddy:
    build: ./dockerfile-caddy
    container_name: caddy
    hostname: caddy
    restart: unless-stopped
    ports:
      - "80:80"
      - "443:443"
    environment:
      - MY_DOMAIN
      - CLOUDFLARE_API_TOKEN
    volumes:
      - ./Caddyfile:/etc/caddy/Caddyfile:ro
      - ./data:/data
      - ./config:/config

networks:
  default:
    name: $DOCKER_MY_NETWORK
    external: true

- Edit Caddyfile

Add global option acme_dns
or add tls directive to the site-blocks.

Caddyfile

{
  acme_dns cloudflare {$CLOUDFLARE_API_TOKEN}
}

a.{$MY_DOMAIN} {
    reverse_proxy whoami:80
}

b.{$MY_DOMAIN} {
    reverse_proxy nginx:80
    tls {
        dns cloudflare {$CLOUDFLARE_API_TOKEN}
    }
}

- Wildcard certificate

A one certificate to rule all subdomains. But not apex/naked domain, thats separate.
As shown in the documentation, the subdomains must be moved under the wildcard site block and make use of host matching and handles.

Caddyfile

{
  acme_dns cloudflare {$CLOUDFLARE_API_TOKEN}
}

{$MY_DOMAIN} {
    reverse_proxy homer:8080
}

*.{$MY_DOMAIN} {
    @a host a.{$MY_DOMAIN}
    handle @a {
        reverse_proxy whoami:80
    }

    @b host b.{$MY_DOMAIN}
    handle @b {
        reverse_proxy nginx:80
    }

    handle {
        abort
    }
}

Here's some discussion on this and a simple, elegant way we could have had, without the need to dick with the Caddyfile this much. Just one global line declaration. But the effort went sideways.
So I myself do not even bother with wildcard when the config ends up looking complex and ugly.