The image's init (migrate.sh) sent the supabase_admin password as SQL text while its postgresql.conf logged DDL, so the password reached the container log. The server now starts with -c log_statement=none, which the entrypoint also hands to the init server; a harness check pins it, and DEPLOY.md makes the post-init log scan a production prerequisite.
35 KiB
Deploy
Status of this guide (2026-10-25)
This guide was written for the static deployment of 0.1.0, the version
faunosteo.com runs, and is kept as that record. It is not a procedure for
develop: nothing authorises deploying it (release readiness track, TODO
RR-1 to RR-10). The corrections this file needs are TODO RR-5.
Historical, true for 0.1.0 and wrong for an image with a backend: the introduction below ("no backend, no database"), "Data persistence", "Backups", "Backend CORS", "Frontend API base URL", the rollback's "no server-side state", and the troubleshooting entries "CORS rejected in browser" and "volume / data unexpectedly empty".
Not verified for the current code:
- the loopback claim:
docker-compose.ymlpublishes${PORT:-8080}on every interface, and rootful Docker bypasses UFW; - "First-time VPS setup", "Routine deploys" and "Rollback" build the image on the VPS; an image from R2 on compiles the Rust core and is never built there (see "Image deploy by digest");
- the fallback that serves
apps/app/dist/from a plainnginx:alpinedropsdocker/nginx.conf(theenv.jsno-cache rule, the.wasm.gzserved withgzip_static); - the VPS sizing table, written before the Rust build stages.
Would break the current application if applied as written:
- the two CSP examples (Option A nginx, Option B Caddy), marked below as not
production-ready: without
'wasm-unsafe-eval'the editor, the export and the import cannot load the Rust core; without the Supabase origin inconnect-src(it falls back todefault-src 'self'), sign-in and the/appspace fail unless Supabase is served from the same origin; - a default build (
SHEET_BUNDLE=none): no anatomical sheet, the editor refuses to create or import a report.
Security prerequisites before deploying anything built from develop:
- the CSP of "WebAssembly and the CSP" with
connect-srcnaming the Supabase origin, checked on the image (docker/csp-proof/app-image.sh); - the container port bound to loopback in fact, checked from outside the VPS;
- a hosting decision for Supabase (
docs/development/OPEN_QUESTIONS.mdQ2), production SMTP with French auth email templates, captcha, OAuth ids, MFA for platform roles (RR-3, RR-5); - the worker's service key out of
.env, its exposure to thedockergroup settled (RR-4); - images built off the VPS and pulled by digest, base images pinned (RR-6);
- a self-hosted Supabase database on the
supabase/postgresimage starts with-c log_statement=noneon its server command line from its very first start (the image's init sends role passwords as SQL text while itspostgresql.conflogs DDL; the entrypoint passes the flag to the init server), and right after init its container log (and log files, if any) is scanned for the password, counts only, before anything else (docs/development/SUPABASE.md§5.2); - the RLS and API items of RR-3 and RR-4 and the dependency advisories of RR-1 treated or qualified.
Single nginx:alpine container behind a host reverse proxy that terminates
TLS. The container serves a static SPA bundle (built from apps/app, Node 24,
pnpm 10.13.1) from /usr/share/nginx/html
and a runtime-injected env.js; there is no backend, no database, no
persistent server-side state. Praticien data lives in the browser
(sessionStorage / localStorage) and in the exported PDFs themselves
(JSON attached natively to each PDF). The container runs entirely as a
non-root nginx user — no process is root after the image is built.
What gets deployed
| Service | Image source | Port (container) | Port (host) | Role |
|---|---|---|---|---|
app |
local build (multi-stage Dockerfile) |
${PORT:-8080} |
${PORT:-8080} |
Serves the SPA (dist/) + runtime env.js. nginx:1.31-alpine inside, runs as non-root nginx user, SPA fallback to index.html. |
The host should publish only ports 22, 80, 443 to the internet. The
container port (8080 by default) should stay loopback-only — the host
reverse proxy reads from it on 127.0.0.1:8080.
Rust (plan R0/R1, docs/development/RUST.md): the repository also builds a
worker image (crates/worker/Dockerfile, static musl binary on
distroless/static, container faunosteo_worker, service worker under the
Compose profile worker). Since Step 8 it carries the files job loop (avatar
processing, purge sweeps: docs/development/STORAGE.md §3); since V1-2.3 the
same image also runs the sheet verification mode, service worker-verify
(run --jobs sheet-verify, STORAGE.md §3.6), never started on DEV without the
owner's go. It is not deployed
yet: it needs the self-hosted Supabase stack beside it, which production does
not run today; the live deployment above is unchanged. The
production VPS never needs a Rust toolchain: Rust is compiled on the Ubuntu
development machine (native toolchain) or inside Docker builder images, and
production images will be pulled by digest ("Image deploy by digest",
prepared below) once a runner and a registry exist on the forge. No CI/CD is
operational today (2026-10-13).
Reference sheet in the image (SHEET_BUNDLE)
The application needs an anatomical sheet to work. Until sheets are served by Supabase (plan Step 13) the only source is the sheet bundled at build time:
| Build | Sheet in the bundle |
|---|---|
docker build . / pnpm build (default, CI) |
none: the app shows "no anatomical sheet" and refuses to create or import a report |
docker build --build-arg SHEET_BUNDLE=draft . / pnpm build:preview |
the curated equine DRAFT, medically unvalidated |
Bundling the draft is a deliberate preview decision, never the default: the draft content is readable by anyone who can load the application (BDR-037).
Docker resource naming
Expected change in .env.example (file not editable by the assistant):
replace COMPOSE_PROJECT_NAME=osteo-report with COMPOSE_PROJECT_NAME=faunosteo
or drop the line. The variable overrides name: faunosteo in
docker-compose.yml, so an .env copied from the example would still create
osteo-report_* resources. No secret is involved.
Several SaaS may share one VPS, so every Docker resource Faunosteo owns
carries the faunosteo_ prefix whenever its name is configurable:
containers, Compose project (hence networks and named volumes), images and
future services (faunosteo_app, faunosteo_worker, faunosteo_supabase_db…).
| Resource | Name | Set by |
|---|---|---|
| SPA container | faunosteo_app |
container_name in docker-compose.yml |
| Compose project, network | faunosteo, faunosteo_default |
top-level name: in docker-compose.yml |
| Local Supabase stack (CLI) | supabase_<service>_faunosteo |
project_id in supabase/config.toml |
Known limit: the Supabase CLI hard-codes supabase_<service>_<project_id>
for its containers, volumes and network (supabase_network_faunosteo); the
project id is a suffix, not a prefix. The name stays unique per SaaS, so no
workaround is applied. A self-hosted production Supabase (docs/development/
SUPABASE.md §10) is a Compose project we write ourselves and must follow the
prefix (faunosteo_supabase_db, faunosteo_supabase_auth, …).
Renaming a container never touches data: the SPA container holds no volume.
When a future service gains a named volume, keep the volume name across a
rename (external: true or an explicit name:) instead of recreating it.
Required environment
All vars live in a single .env file at the project root on the VPS
(read by docker compose via env_file: .env). Defaults sit in
docker/entrypoint.sh and src/config/env.ts; production should set
each one explicitly.
| Variable | Required | Default | Notes |
|---|---|---|---|
APP_SITE_NAME |
yes | Faunosteo |
Browser <title>, OG site_name. |
APP_BRAND_NAME |
yes | Faunosteo |
Header brand block, PDF header. Praticien-visible identity. |
APP_FOOTER_NAME |
yes | Faunosteo |
Footer copyright line. |
APP_TAGLINE |
yes | Outils pour ostéopathes animaliers |
Hero tagline on Home. |
APP_SUPABASE_URL |
no | empty | Supabase project URL exposed to the browser. Empty = app runs without backend (current mode). |
APP_SUPABASE_ANON_KEY |
no | empty | Supabase anon (publishable) key. Public by design, data stays behind RLS. Never the service_role key. |
APP_AUTH_PROVIDERS |
no | empty | OAuth buttons on the auth screens, comma-separated (google,apple). The Supabase project must enable the same providers; empty = email/password only. |
PORT |
no | 8080 |
Container listen port. Must be >=1024 — the container runs as a non-root nginx user and cannot bind privileged ports. |
No secrets are required in this file: the anon key is public by design and
the service_role key never leaves the Supabase host (see
docs/development/SUPABASE.md).
The .env file is not sensitive but should still be chmod 600 to keep
local identity overrides private.
Worker environment (.env.worker, Step 8)
The files worker (docker compose --profile worker up -d worker) reads a
second env file, .env.worker, never the app .env: it holds the service
key. chmod 600, excluded from every backup copy of env files, readable
through docker inspect by the docker group (documented limit). The
assistant cannot write .env* files in this repository: create it by hand
from the table (STORAGE.md §3.2 lists the optional tuning variables).
| Variable | Required | Notes |
|---|---|---|
SUPABASE_URL |
yes | Kong origin on the Supabase compose network (http://supabase_kong_faunosteo:8000 on the local CLI stack; the self-hosted stack's own Kong name in production). Never the public URL when the stack is on the same host. |
SUPABASE_SERVICE_ROLE_KEY |
yes | the service key; on self-hosted a long-lived JWT (revoking it = rotating JWT_SECRET). Never in .env, env.js or the repository. |
WORKER_ID |
no | tag of the log lines (default worker-<pid>). |
The service joins the external network named by SUPABASE_NETWORK
(default supabase_network_faunosteo, the local CLI stack); it publishes no
port, runs read-only with every capability dropped, 512 MiB, 1 CPU, 64 pids.
The verification service worker-verify (V1-2.3) has the same hardening with
1 280 MiB, so its child's 1 GiB address-space limit binds first: a host that
runs it beside the stack needs that memory on top of the 1 GB below.
Browser side, the production CSP connect-src must name the Supabase
origin the app talks to (already required since Step 7), and img-src
keeps blob: (the avatar is shown from a blob URL downloaded under RLS).
Provision the VPS
| Resource | Recommended minimum |
|---|---|
| CPU | 1 vCPU |
| RAM | 1 GB (build runs out-of-band — see below) |
| Disk | 10 GB |
| OS | Ubuntu 24.04 LTS (or any recent Debian-family) |
Build-time RAM peaks around 1 GB (Vite + TS + Tailwind). If your VPS only has 1 GB, either swap-enable the host or build the image locally and push it to a registry — the runtime image is well under 100 MB.
DNS:
# A record on your domain registrar / DNS provider
faunosteo.com → <vps-ipv4>
faunosteo.com (AAAA) → <vps-ipv6> # optional
SSH key copied to the VPS before disabling password auth (see Hardening below).
Baseline packages (apt):
apt-get update
apt-get install -y git curl ufw fail2ban ca-certificates gnupg \
nginx certbot python3-certbot-nginx unattended-upgrades
Firewall (two layers)
Layer 1 — cloud provider security group
| Source | Port | Protocol | Rule |
|---|---|---|---|
| your office / home IP | 22 | TCP | allow |
| 0.0.0.0/0 | 80 | TCP | allow |
| 0.0.0.0/0 | 443 | TCP | allow |
| any | other | any | deny |
Cheat sheet:
- Scaleway: Console → Security Groups → create + bind to instance.
- Hetzner: Console → Firewalls → attach to server.
- OVH: Public Cloud → Networking → Security groups (legacy: iptables).
- DigitalOcean: Networking → Firewalls.
Layer 2 — UFW on the VM
ufw default deny incoming
ufw default allow outgoing
ufw allow 22/tcp # SSH
ufw allow 80/tcp # HTTP (for ACME challenge + redirect)
ufw allow 443/tcp # HTTPS
ufw --force enable
ufw status verbose
Do not open 8080 (or whatever PORT you pick). The host nginx talks
to the container on 127.0.0.1:8080.
Install Docker (production-tuned)
curl -fsSL https://get.docker.com | sh
systemctl enable --now docker
# Cap container log volume — JSON-file driver with rotation
cat > /etc/docker/daemon.json <<'EOF'
{
"log-driver": "json-file",
"log-opts": {
"max-size": "10m",
"max-file": "5"
},
"live-restore": true
}
EOF
systemctl restart docker
Confirm:
docker version
docker info | grep -i 'logging driver\|live restore'
First-time VPS setup
# As your non-root deploy user (see Hardening §1 if you have not made one yet)
cd ~
git clone https://git.bchanot.fr/bchanot/faunosteo.git
cd faunosteo
# Configure identity
cp .env.example .env
$EDITOR .env # set APP_BRAND_NAME, APP_FOOTER_NAME, APP_TAGLINE, etc.
# First build (publishes to 127.0.0.1:8080)
docker compose up --build -d
docker compose ps
docker compose logs --tail=50
# Loopback sanity check
curl -sI http://127.0.0.1:8080/ | head -1 # expect 200 OK
curl -s http://127.0.0.1:8080/env.js # expect window.__APP_CONFIG__ = {...};
If the build OOMs on a 1 GB VPS, build locally (pnpm build) and rsync the
resulting apps/app/dist/ over, then bind-mount it into a plain nginx:alpine until you
upsize.
Routine deploys
This project has no scripts/deploy.sh yet. Pull-and-recompose from the
VPS is the path:
cd ~/faunosteo
git fetch origin
git reset --hard origin/main # default branch is "main" on this remote
docker compose up --build -d
docker image prune -f # reclaim old build layers
Verify post-deploy:
docker compose ps
curl -sI https://faunosteo.com/ | head -3
For zero-downtime updates, docker compose up --build -d rebuilds the
image then recreates the container; the container restart is sub-second
because nginx:alpine boots instantly. There is no warmup phase.
Rollback note (PDF archives, Step 6): PDFs exported by a Step 6 build carry a report archive envelope. A build older than Step 6 reads such a PDF as "no attachment"; a Step 6 build reads every older PDF (bare v2 and v3). Roll back only with that limit in mind and tell the practitioners which build produced their files.
Rollback note (PDF archives, R4): PDFs written by an R4 build (the core's
writer) are read by the R3 build and by release 0.2.0: the frozen copy of
their reader (packages/core-wasm/tests/n-1/scanner, O11) and the frozen
Zod schemas (n-1/domain) read every R4 fixture in the test suites
(n-1.reader, n-1.scanner, archive.envelopes); the attachment bytes are
the envelope the TypeScript wrote, in contract key order (D-R11). An R4
build reads every older PDF (bare v2, bare v3, Step 6 and R3 envelopes).
The TypeScript PDF engine left the repository at R4.8: a rollback is a
previous image, never a rebuild of the old engine.
Data persistence
Historical (0.1.0, no backend): once Supabase or the worker run beside the application, the database and the Storage objects are server-side state that this section and the next do not cover.
There are no server-side volumes to back up:
- In-flight session data lives in the praticien's browser
(
sessionStorage osteo:draft,localStorage osteo:lastPractitioner,localStorage osteo:locale,sessionStorage osteo:disclaimer-shown). - Long-term record is the exported PDF — the full report JSON is
attached natively (
AFRelationship.Data). Re-importing the PDF restores the session.
The only "data" the server holds is the SPA bundle and env.js, both
regenerated on every build.
Backups (cron + retention)
Historical (0.1.0): the .env snapshot below is the whole backup only while
the server holds no data; docs/development/STORAGE.md §5.2 sets the order
for a backend (database dump first, files within 48 h).
Because there is no server-side state, the backup target is the source
repository plus the .env file.
.env snapshot
# /opt/backup-osteo.sh
#!/bin/sh
set -e
DEST="/var/backups/osteo"
DATE="$(date -u +%Y%m%d)"
mkdir -p "$DEST"
cp /home/<deploy-user>/faunosteo/.env "$DEST/env-$DATE"
chmod 600 "$DEST/env-$DATE"
find "$DEST" -name 'env-*' -mtime +30 -delete
Make executable:
chmod 700 /opt/backup-osteo.sh
chown root:root /opt/backup-osteo.sh
Cron
# crontab -e (root)
5 3 * * * /opt/backup-osteo.sh >> /var/log/backup-osteo.log 2>&1
Off-VPS storage (optional but recommended)
Sync the backup directory to S3-compatible storage (Scaleway Object Storage, B2, etc.):
# After installing rclone or aws-cli
aws s3 sync /var/backups/osteo s3://<bucket>/osteo/ --storage-class STANDARD_IA
Restore
docker compose down
cp /var/backups/osteo/env-<date> /home/<deploy-user>/faunosteo/.env
docker compose up --build -d
curl -sI https://faunosteo.com/
The praticien's exported PDFs are the real long-term store. If the VPS
disappeared entirely, a fresh git clone + .env + docker compose up
on a new host reproduces the live app.
Behind a domain + TLS reverse proxy
Option A — host nginx + Certbot
DNS A/AAAA records pointing at the VPS, UFW allowing 80/443, then on the VPS:
# /etc/nginx/sites-available/osteo
server {
listen 80;
listen [::]:80;
server_name faunosteo.com;
# Certbot will rewrite this block to add the 301 redirect.
location /.well-known/acme-challenge/ {
root /var/www/letsencrypt;
}
location / {
proxy_pass http://127.0.0.1:8080;
proxy_http_version 1.1;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
}
}
server {
# nginx >= 1.25.1: HTTP/2 is its own directive (`listen ... http2` is deprecated).
# nginx 1.24 (Ubuntu 24.04 package) does not know `http2 on;`: there, write
# `listen 443 ssl http2;` and `listen [::]:443 ssl http2;` instead. Check with `nginx -v`.
listen 443 ssl;
listen [::]:443 ssl;
http2 on;
server_name faunosteo.com;
# ssl_certificate + ssl_certificate_key written by certbot
location / {
proxy_pass http://127.0.0.1:8080;
proxy_http_version 1.1;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
}
}
Enable + obtain certificate:
ln -s /etc/nginx/sites-available/osteo /etc/nginx/sites-enabled/osteo
nginx -t && systemctl reload nginx
certbot --nginx -d faunosteo.com
systemctl status certbot.timer # auto-renew preinstalled on Ubuntu
Security headers (Option A only)
Not production-ready (2026-10-25): the Content-Security-Policy line below is
the 0.1.0 header. It lacks 'wasm-unsafe-eval' and the Supabase origin in
connect-src; any later image needs "WebAssembly and the CSP" below first
(correction tracked as TODO RR-5).
Add inside the listen 443 block (under any existing location rules
the certbot template inserted):
add_header Strict-Transport-Security "max-age=63072000; includeSubDomains; preload" always;
add_header X-Content-Type-Options "nosniff" always;
add_header X-Frame-Options "DENY" always;
add_header Referrer-Policy "strict-origin-when-cross-origin" always;
add_header Permissions-Policy "camera=(), microphone=(), geolocation=()" always;
add_header Content-Security-Policy "default-src 'self'; img-src 'self' data: blob:; style-src 'self' 'unsafe-inline'; script-src 'self'; font-src 'self' data:; object-src 'none'; base-uri 'self'; frame-ancestors 'none'" always;
Reload + verify:
nginx -t && systemctl reload nginx
curl -sI https://faunosteo.com/ | grep -iE 'strict-transport|x-frame|content-security'
# Cross-check externally:
# https://securityheaders.com/?q=faunosteo.com
# https://www.ssllabs.com/ssltest/analyze.html?d=faunosteo.com
The 'unsafe-inline' on style-src is needed because Tailwind injects
inline <style> tags via the runtime engine; tighten if a Tailwind
build-time export gets adopted later. img-src includes data: and
blob: because the canvas rasterization pipeline produces blob URLs and
the PDF preview emits data URLs.
Option B — Caddy (auto-TLS, single file)
Not production-ready (2026-10-25): the Content-Security-Policy of this
Caddyfile is the 0.1.0 header, with the same gaps as Option A (TODO RR-5).
apt-get install -y debian-keyring debian-archive-keyring apt-transport-https
curl -1sLf https://dl.cloudsmith.io/public/caddy/stable/gpg.key | gpg --dearmor -o /usr/share/keyrings/caddy-stable-archive-keyring.gpg
echo "deb [signed-by=/usr/share/keyrings/caddy-stable-archive-keyring.gpg] https://dl.cloudsmith.io/public/caddy/stable/deb/debian any-version main" \
> /etc/apt/sources.list.d/caddy-stable.list
apt-get update && apt-get install -y caddy
cat > /etc/caddy/Caddyfile <<'EOF'
faunosteo.com {
encode gzip zstd
reverse_proxy 127.0.0.1:8080
header {
Strict-Transport-Security "max-age=63072000; includeSubDomains; preload"
X-Content-Type-Options "nosniff"
X-Frame-Options "DENY"
Referrer-Policy "strict-origin-when-cross-origin"
Permissions-Policy "camera=(), microphone=(), geolocation=()"
Content-Security-Policy "default-src 'self'; img-src 'self' data: blob:; style-src 'self' 'unsafe-inline'; script-src 'self'; font-src 'self' data:; object-src 'none'; base-uri 'self'; frame-ancestors 'none'"
}
}
EOF
systemctl reload caddy
Caddy handles TLS issuance + renewal automatically.
Backend CORS
Historical (0.1.0): since Step 7 the browser calls the Supabase API, a cross-origin request unless Supabase is served from the same origin.
N/A. The app has no backend and makes no cross-origin requests. The only
runtime fetches target the same origin (SVG assets, env.js).
Frontend API base URL
N/A. There is no build-time API URL var; the app is static.
Sanity checks
# HTTPS redirect on port 80
curl -sI http://faunosteo.com/ | grep -i location # expect Location: https://...
# TLS itself
curl -sI https://faunosteo.com/ | head -3 # expect HTTP/2 200
# env.js shipped fresh (no caching)
curl -sI https://faunosteo.com/env.js | grep -i cache-control
# expect: Cache-Control: no-store, no-cache, must-revalidate
# Internal port is NOT publicly bound
nc -zv <vps-public-ip> 8080 2>&1 | grep -i refused # expect "Connection refused"
Post-deploy hardening
1. Non-root deploy user
adduser deploy
usermod -aG docker deploy
mkdir -p /home/deploy/.ssh
cp ~/.ssh/authorized_keys /home/deploy/.ssh/
chown -R deploy:deploy /home/deploy/.ssh
chmod 700 /home/deploy/.ssh
chmod 600 /home/deploy/.ssh/authorized_keys
echo 'deploy ALL=(ALL) NOPASSWD: /usr/bin/systemctl, /usr/bin/journalctl' \
> /etc/sudoers.d/deploy
2. Disable root + password SSH
In /etc/ssh/sshd_config:
PermitRootLogin prohibit-password
PasswordAuthentication no
PubkeyAuthentication yes
Reload + verify on a second terminal before closing the first one:
systemctl reload ssh
ssh deploy@faunosteo.com 'echo ok'
3. (Optional) Move SSH off port 22
Open the new port everywhere before closing 22:
ufw allow 2222/tcp # 1. UFW
# 2. Add 2222 to cloud security group
# 3. /etc/ssh/sshd_config: Port 2222 (keep Port 22 temporarily)
systemctl reload ssh
ssh -p 2222 deploy@faunosteo.com 'echo ok' # confirm
# 4. Remove Port 22 from sshd_config + UFW + cloud SG
4. Unattended security upgrades
apt-get install -y unattended-upgrades
dpkg-reconfigure --priority=low unattended-upgrades
Confirm /etc/apt/apt.conf.d/20auto-upgrades enables Update-Package-Lists
and Unattended-Upgrade.
5. fail2ban
systemctl enable --now fail2ban
fail2ban-client status sshd
6. Final lockdown check
# From OUTSIDE the VPS
nmap -Pn -p 22,80,443,8080,2222 <vps-public-ip>
# 22: filtered (if SSH moved) or open (if not)
# 80: open
# 443: open
# 8080: filtered ← MUST NOT be open
# 2222: open ← if SSH was moved
# Password-SSH attempt — should fail
ssh -o PreferredAuthentications=password -o PubkeyAuthentication=no deploy@faunosteo.com
# Expect: Permission denied (publickey).
Rollback
cd ~/faunosteo
git reflog | head -20 # find the known-good SHA
git reset --hard <known-good-sha>
docker compose up --build -d
curl -sI https://faunosteo.com/ | head -1
Because there is no server-side state to corrupt, rollback is purely a code revert + rebuild. Praticien browsers retain their own data; PDFs already exported are unaffected.
Image deploy by digest (prepared at R0, not yet applied)
Decision D-R13 (docs/development/RUST_ARCHITECTURE_REVIEW.md Annex A and
C): production images are built in CI and the VPS pulls them; it never
compiles the Rust project, and the DEV1-XS memory limit stops mattering for
builds. The procedure below is prepared and documented now; it replaces
"Routine deploys" and "Rollback" above only when the first image built
this way is deployed, after a separate validation. Until then the live
deployment keeps building on the VPS as documented above.
From R2 on (feature/rust-r2-domain-core), the application image
compiles the Rust core (Dockerfile stage wasm: the Rust toolchain and
wasm-bindgen-cli in a builder stage). Such an image is never built on the
VPS (1 vCPU, 1 GB): it is built on the development machine or a runner and
pulled by digest, with the CSP of the next section in place first. The live
deployment, built from main before R2, is not affected.
- Build and publish: run the manual workflow
release-images(.gitea/workflows/release-images.yml,workflow_dispatch, never on push) on the forge. Prerequisite not met yet: no runner, registry or secret is configured on the forge (2026-10-13); the workflow has never run, which is the CI/CD step to decide separately. It pushesfaunosteo_app:<sha>andfaunosteo_worker:<sha>to the container registry named by theREGISTRYsecret (the forge's own Gitea registry is the recommended choice, Scaleway Container Registry the alternative) and prints their digests in the job summary. No workflow deploys anything. - Pin:
docker-compose.ymlon the VPS references the images by digest (image: <registry>/<owner>/faunosteo_app@sha256:…); the digest is the only thing that changes at a deploy, by hand, after validation. - Deploy:
docker compose pull && docker compose up -d(no--build), then the post-deploy checks of "Routine deploys". - Rollback: set the previous digest back in
docker-compose.ymland run step 3 again; nothing is rebuilt and nothing needs crates.io or npm. - Registry credentials on the VPS: a read-only token in
~/.docker/config.jsonof the deploy user, never in the repository.
Prerequisites checked at R0 and to re-check before the first digest deploy:
disk headroom on the VPS for two image generations (app ≈ 100 MB,
worker ≈ 9 MB, measured at R0), the registry reachable from the VPS, docker compose pull
working with the read-only token.
WebAssembly and the CSP (from R2 on)
From R2 the application loads the Rust core module in the sheet pages (the
editor, the new-report submit, the PDF import). Before an R2 image is
deployed, the two Content-Security-Policy headers of this file (nginx and
Caddy options above) must carry 'wasm-unsafe-eval' in script-src; the
examples above still show the header of the live, pre-R2 deployment:
script-src 'self' 'wasm-unsafe-eval';
Without the keyword an R2 image stays usable but degraded: the sign-in and
account pages work, the editor shows "Le moteur de calcul de Faunosteo n'a
pas pu être chargé" with a retry, and the tab's draft is left untouched
(proven on the image by docker/csp-proof/app-image.sh,
docs/development/RUST.md §8.5). The keyword adds nothing for a pre-R2
image, so it can be set on the proxy before the switch.
Proven at R0 under the production nginx image and headers with headless
Chrome for Testing 156 (docker/csp-proof/run.sh, result in
docs/development/RUST.md §5): with script-src 'self' the browser refuses
to compile the module ("Compiling or instantiating WebAssembly module
violates the following Content Security policy directive"); with the keyword
the module instantiates. 'wasm-unsafe-eval' allows WebAssembly compilation
only, never eval or inline scripts (the proof's inline variant was still
blocked). Browser floor for the keyword: Chrome 97, Firefox 102, Safari 16
(the Vite 7 baseline). The .wasm file is served as application/wasm by
nginx:1.31-alpine out of the box; since R2 the image precompresses it and
docker/nginx.conf serves the .wasm.gz with gzip_static on; (R4.8
module with the PDF reader, writer and session: 1,535,289 B raw, 519,795 B
gzip -9; R3: 1.18 MB raw, 372 KB served; D-R13 budget 600 KB, report line
550 KB passed at R4.6 and declared). A proxy in front must pass
Content-Encoding through and keep the application/wasm type.
Since R4 the PDF import runs in a module worker (/assets/import.worker-*.js,
a same-origin hashed file that script-src 'self' already allows; no
blob:, no inline worker): the worker's response carries the same policy, so
the keyword is required for the import too. Without it the import answers
« Le moteur de lecture n'a pas pu être chargé » and nothing else changes. The
CSP needs no other change. Browser support of the import, apart from the
application's floor above: module workers need Firefox 114 or later (Chrome
80, Safari 15); the import already needed Firefox 113 and Safari 16.4 before
R4 (DecompressionStream), so Firefox 113 is the one version that loses it
(owner decision OD-1, docs/development/RUST.md §10). The worker never
fetches the module when the page's runtime already holds it (the compiled
module is posted), so an import works offline once the editor has loaded
the engine, as before.
Health checks
| URL | Expected |
|---|---|
https://faunosteo.com/ |
HTTP/2 200, HTML containing <div id="root"> |
https://faunosteo.com/env.js |
HTTP/2 200, Cache-Control: no-store, no-cache, must-revalidate, body starts with window.__APP_CONFIG__ = { |
https://faunosteo.com/assets/<hash>.js |
HTTP/2 200, Cache-Control: public, max-age=31536000, immutable |
https://faunosteo.com/nonexistent |
HTTP/2 200 (SPA fallback serves index.html) |
http://127.0.0.1:8080/ (on VPS) |
HTTP/1.1 200 |
There is no /health endpoint — the container only serves static assets.
Liveness is "nginx process up + serving 200 on /".
Troubleshooting
container won't start
docker compose logs --tail=200 app
Common: .env missing or unreadable (compose will still start but
APP_* defaults apply); permission error on /entrypoint.sh (rebuild
with --no-cache).
5xx from the site
The container itself only serves static files — a 5xx almost always comes from the host reverse proxy.
journalctl -u nginx --since "10 min ago" # Option A
journalctl -u caddy --since "10 min ago" # Option B
Most common: proxy_pass target wrong (127.0.0.1:8080 vs container
port) or container down.
CORS rejected in browser
The app makes no cross-origin requests. If you see a CORS error, the
likely cause is a misconfigured CSP — check the Content-Security-Policy
header against the failing resource URL.
502 bad gateway from host nginx
The container is down or listening on a different port. Verify:
docker compose ps
ss -ltnp | grep 8080 # container should own the port
curl -sI http://127.0.0.1:8080/ | head -1
SSL renewal failed
systemctl status certbot.timer
certbot renew --dry-run
journalctl -u certbot --since "1 week ago"
Common: UFW dropping port 80 (ACME HTTP-01 challenge fails) or the host nginx site config moved without reloading.
volume / data unexpectedly empty
N/A — this app has no volumes.
disk full
df -h
docker system df
docker system prune -a -f # purge unused images/layers
journalctl --disk-usage
journalctl --vacuum-time=7d
The Docker log driver is capped at 5×10MB per container in this setup
(see /etc/docker/daemon.json); runaway logs should not happen.
restart loop
docker compose ps # status column
docker compose logs --tail=200 app
Most common: entrypoint.sh non-executable (chmod +x line missing from
Dockerfile after a copy edit) or nginx config syntax error (test with
docker compose run --rm app nginx -t).
container log explosion
Already mitigated by json-file driver caps. If you see disk pressure
from /var/lib/docker/containers/<id>/*.log anyway, lower the cap in
/etc/docker/daemon.json and restart Docker.
Operational notes
- Logs: Docker JSON-file driver, capped at
5 × 10MBper container (/etc/docker/daemon.json). Host nginx logs in/var/log/nginx/; rotate via logrotate (default Ubuntu config). Certbot logs in/var/log/letsencrypt/. - Manual log prune:
docker compose logs --no-color > /tmp/x; truncate -s 0 $(docker inspect --format='{{.LogPath}}' faunosteo_app). - Key rotation:
.envcarries no secret material, but ifAPP_BRAND_NAMEetc. need to change, edit.env+docker compose up -d(no rebuild —entrypoint.shreads env on container start). - Provider-specific:
- Scaleway DEV1-S works; DEV1-XS hits build OOM.
- Hetzner CX22 is the cheapest reliable choice for the build itself.
- On any 1 GB VPS, build locally and skip in-place compose builds.