Both refuse to run unless docker/.env names a non-prod COMPOSE_PROJECT_NAME and scope every compose call to that project. setup.test.sh: like setup.sh but runs the DB/cache/console steps as www-data and repairs var/cache|log|sessions ownership at the end, so php-fpm can read its own compiled cache (bare `docker exec` runs as root -> silent 500s). Never touches var/volumes/db. Test-appropriate final message (NPM upstream, local curl check). restart.test.sh: keeps the database by default (--fresh-db to wipe + re-init), never runs a host-wide docker prune, and only chowns var/cache|log|sessions (chowning var/volumes/db is what corrupted the MySQL data dir). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
6.5 KiB
Test / staging environment (test.escapepage.com)
Runs a second copy of the full Docker stack on the same server as production,
behind the same Nginx Proxy Manager (NPM). Production is untouched: with no
docker/.env overrides, compose.yaml behaves exactly as before.
How isolation works
docker/compose.yaml derives everything instance-specific from docker/.env:
| Concern | Mechanism | prod (no overrides) | test |
|---|---|---|---|
| Container names | ${STACK_NAME:-escapepage}-* |
escapepage-php … |
escapepage-test-php … |
| Compose project (networks, labels) | COMPOSE_PROJECT_NAME |
docker (unchanged) |
escapepage-test |
| Published ports | ${NGINX_HTTP_PORT:-8080} etc. |
0.0.0.0:8080/8443/3306/8090/8025 |
127.0.0.1:8081/8444/3307/8091/8026 |
| Database files | bind mount ../var/volumes/db |
per checkout | per checkout |
| Web reachability | nginx joined to external nginx_default |
via NPM | via NPM |
STACK_NAME is a plain variable (not the Compose-managed COMPOSE_PROJECT_NAME),
so its :-escapepage default is reliable and production needs no docker/.env
change to keep its escapepage-* container names.
nginx, mercure and mailer all sit on the external nginx_default network,
so NPM forwards straight to escapepage-test-nginx / -mercure by name — no
public host port needed.
Production checkout — optional
Nothing is required. Two optional tidy-ups, each needing a docker compose up -d
to recreate the affected container:
- The
nginxservice now also attaches tonginx_default. If you'd rather have NPM forward toescapepage-nginxby name instead ofhost:8080, recreate it and repoint the NPM proxy host. Otherwise the published8080/8443still work as before and you can ignore this. - Add
STACK_NAME=escapepageandCOMPOSE_PROJECT_NAME=escapepageto the productiondocker/.envto move it off the implicitdockerproject name. Cosmetic; do it only if you want the two stacks named symmetrically.
Standing up the test stack
1. DNS
test.escapepage.com → server IP. (mercure-test.escapepage.com too if you want
live game features.) On Cloudflare, DNS-only ("grey cloud") keeps it low-profile.
2. Separate checkout
git clone <repo> /opt/escapepage-test
cd /opt/escapepage-test
git checkout <branch-to-test>
Use a separate clone, not git worktree — the Docker build context is the
checkout directory and full isolation avoids surprises.
3. Compose env
cp docker/.env.test.example docker/.env
# edit: real passwords, fresh MERCURE_JWT_SECRET (openssl rand -hex 32), reCAPTCHA keys
4. Symfony env overrides
Create /opt/escapepage-test/.env.local in the checkout (git-ignored):
APP_SECRET=<openssl rand -hex 16>
# Behind NPM the forwarded client IP lands from a Docker-private range; trust them
# all on staging so URL generation / HTTPS detection work.
TRUSTED_PROXIES=127.0.0.1,10.0.0.0/8,172.16.0.0/12,192.168.0.0/16
APP_ENV=prod, DATABASE_URL, SITE_BASE_URL, MERCURE_* and MAILER_DSN are
already supplied to the containers by docker/.env.
5. Build & start
./docker/setup.test.sh
The test-specific script (not setup.sh): it refuses to run unless
COMPOSE_PROJECT_NAME in docker/.env is a non-prod value, scopes every compose
call to that project, runs the DB/cache/console steps as www-data, and repairs
var/cache / var/log / var/sessions ownership at the end (bare docker exec
runs as root, which otherwise leaves php-fpm unable to read its own cache — a
silent 500). It never touches var/volumes/db.
Builds escapepage-test-* images, starts the containers, creates + migrates
escapepage_test, builds assets. Re-run any time; --no-build skips the rebuild.
6. Nginx Proxy Manager — proxy host
- Domain:
test.escapepage.com - Forward to:
escapepage-test-nginx, port 443, scheme https (the app nginx force-redirects 80→443 and serves a self-signed cert; leave "Verify SSL" off — same arrangement as production) - SSL: request a Let's Encrypt cert, Force SSL, HTTP/2
- Optional: add response header
X-Robots-Tag: noindex
If you want live game screens, add a second proxy host
mercure-test.escapepage.com → escapepage-test-mercure port 80 (http).
7. Restrict access to your IP — NPM Access List
A host firewall on 80/443 can't help: prod and test share those ports on NPM. Restrict at the proxy instead.
- NPM → Access Lists → Add Access List
- Name: e.g.
staging-allowlist - Authorisation: leave empty
- Access:
Allow <your.public.ip>then a finalDeny all - Satisfy: Any
- Name: e.g.
- Edit the
test.escapepage.comproxy host → Access List → select it. - Do the same on
mercure-test.escapepage.comif you added it.
Everyone else gets 403 at the proxy. Update the IP in one place when it changes.
Defence in depth (optional): in /opt/escapepage-test/docker/nginx/default.conf,
inside the server { listen 443 ... } block:
set_real_ip_from 172.16.0.0/12; # NPM's docker network
real_ip_header X-Forwarded-For;
allow <your.public.ip>;
deny all;
Deploying a new version to test
cd /var/sites/escapepage-test
git fetch && git checkout <branch> && git pull
./docker/setup.test.sh --no-build # composer install, migrate, build assets, fix perms
--no-build skips the image rebuild; drop it if the Dockerfile changed.
Restarting
./docker/restart.test.sh # down + up, keeps the DB and images
./docker/restart.test.sh --build # also rebuild images
./docker/restart.test.sh --fresh-db # also wipe var/volumes/db and re-init MySQL
Safety notes
- Use the
*.test.shscripts on this checkout, notsetup.sh/restart.sh. Both refuse to run unlessdocker/.envnames a non-prodCOMPOSE_PROJECT_NAME, and both scope every action to that project — they can't reach the production stack. restart.test.shkeeps the database by default (you loaded prod data into it);--fresh-dbis the only thing that wipes it. It never runs a host-widedocker system prune/docker builder prune, and it only repairs ownership ofvar/cache/var/log/var/sessions— nevervar/volumes/db(chowning the MySQL data dir is what corrupted it earlier this build).- The test
docker/.envuses its ownDB_NAMEand passwords so a config slip can't reach the production database. MAILER_DSN=smtp://mailer:1026keeps staging mail inside Mailpit instead of sending through Mailgun.