Adds the stop-Phoenix drill to the runbook, using forced probe runs so the outage stays under a minute, and a dated drill log with the first run's results. The log notes that the manual probe was started twice during the outage, which is why two notifications arrived for one down/up transition.
49 KiB
Production deployment
Phoenix release on the shared Ubuntu VPS, behind Caddy, with Postgres in Docker.
Internet
└─ Cloudflare (TLS, WAF, edge rate limiting)
└─ Caddy :443 (cloudflare ranges only)
└─ Phoenix release 127.0.0.1:4100
└─ Postgres 127.0.0.1:5432
Pinned versions
| Thing | Version |
|---|---|
| OS | Ubuntu 24.04.5 LTS, x86_64 |
| glibc | 2.39 |
| Erlang | 28.5 |
| Elixir | 1.19.5-otp-28 |
| Rust | 1.98.1 |
| Node.js | 24 (asset build only) |
| Postgres | postgres:18.6, pinned directly in the unified compose |
| Caddy | 2.11.9 |
| Docker | 29.9.3 |
| Docker Compose | v5.7.0 |
| GitHub runner | ubuntu-24.04 |
The release ships its own ERTS. The host needs the runtime libraries for the
Rust NIF (libstdc++6), OpenSSL, and CA certificates.
Host prerequisites
A new Ubuntu 24.04 VPS needs:
- Docker Engine and the Compose plugin
- Caddy
curl,ca-certificates,libstdc++6,openssl- Node.js 24 + npm if you build the release on this host instead of CI
- DNS and TLS for the domain
- A non-root deploy user with SSH access
Install the host libraries:
sudo apt update
sudo apt install -y curl ca-certificates libstdc++6 openssl
Install Docker and Caddy from their official repositories. The Caddy install
commands already used for the static site are in deploy/static/README.md.
Firewall:
sudo ufw allow OpenSSH
sudo ufw allow 80/tcp
sudo ufw allow 443/tcp
sudo ufw enable
Do not open 4100 or 5432. Phoenix binds to loopback, and Postgres is
published only on 127.0.0.1.
DNS:
A/AAAAforinnersquare.appandwwwpoint at the VPS.- Do not remove the existing mail records:
- root MX records for
support@innersquare.app - SES DKIM CNAMEs
- the
mail.innersquare.appcustom MAIL FROM MX/TXT records - DMARC at
_dmarc.innersquare.app
- root MX records for
Shared-host choices
- Port 4100 is used for Phoenix because another BEAM process already owns
4000on this VPS. - Postgres runs as
chesstrainer-dbin the VPS's unifieddocker-compose.ymlon127.0.0.1:5432; that port is free. The image version is pinned directly in that file. - Caddy is shared and managed manually. This repo ships a reference block
in
deploy/Caddyfile; it is never installed automatically. - The Postgres volume is named
chesstrainer_pgdata, so it cannot collide with other Docker projects.
Directory layout
/opt/chesstrainer/
├── releases/
│ ├── 20261109120000/
│ └── 20261109153000/
├── current -> releases/20261109153000
└── src/ # optional checkout of this repo / deploy scripts
/etc/chesstrainer/env is the application environment file, owned by
root:root with mode 0600.
/etc/chesstrainer/postgres.env holds the Postgres container credentials,
also root:root and 0600.
/etc/chesstrainer/probe.env holds the public probe's URL and its
healthchecks.io ping URL, also root:root and 0600. See §Monitoring and
alerting.
The app runs as the dedicated chesstrainer system user. It must own
/opt/chesstrainer so it can write priv/static/backups.
Database dumps live in /var/backups/chesstrainer, outside /opt/chesstrainer
and priv/static. See §Backups.
Copy the deploy artifacts
Run these commands from a checkout of this repo on the VPS, or copy the
deploy/ directory to /opt/chesstrainer/src/deploy.
sudo install -d -o root -g root /opt/chesstrainer/src/deploy
sudo cp -a deploy/. /opt/chesstrainer/src/deploy/
If you prefer to keep the checkout elsewhere, use that path in the commands below.
Application env file
/etc/chesstrainer/env:
SECRET_KEY_BASE=...
DATABASE_URL=ecto://chesstrainer:...@127.0.0.1:5432/chesstrainer_prod
PHX_HOST=innersquare.app
PORT=4100
POOL_SIZE=10
TRUST_FORWARDED_FOR=true
PRELAUNCH=true
MAILING_LIST_ENABLED=false
SES_REGION=eu-north-1
SES_ACCESS_KEY_ID=AKIA...
SES_SECRET_ACCESS_KEY=...
MATOMO_URL=https://analytics.owenrees.eu/
MATOMO_SITE_ID=6
TABLEBASE_PATH=/opt/chesstrainer/tablebases
# Optional:
# RELEASE_COOKIE=...
The production endpoint binds to 127.0.0.1 only. Caddy is the public
entry point.
Matomo analytics
Set MATOMO_URL and MATOMO_SITE_ID to enable cookieless analytics on public
pages and the guest trainer. Signed-in account and admin pages are excluded.
The privacy controls, Matomo server settings, retention, and opt-out behaviour
are documented in docs/analytics.md.
Generate the two secrets:
openssl rand -base64 64 # SECRET_KEY_BASE
openssl rand -hex 32 # Postgres password
Secret key base
SECRET_KEY_BASE is required in production; the release raises at boot without
it. It signs cookies, session tokens, LiveView tokens, and mailing-list
confirmation links.
Generate it with mix phx.gen.secret or:
openssl rand -base64 64
Keep it stable. Rotating it signs every user out and invalidates outstanding confirmation and unsubscribe links. Never reuse the dev/test value or commit it.
Verify it is set without printing it:
sudo awk -F= '/^SECRET_KEY_BASE=/{print length($2)}' /etc/chesstrainer/env
Postgres env file
/etc/chesstrainer/postgres.env:
POSTGRES_USER=chesstrainer
POSTGRES_PASSWORD=...
POSTGRES_DB=chesstrainer_prod
Postgres 18 stores data under /var/lib/postgresql/18/docker, so the
Compose volume mounts the parent:
volumes:
- chesstrainer_pgdata:/var/lib/postgresql
Local tablebases (optional)
The production tablebase directory is /opt/chesstrainer/tablebases. Keep the
.rtbw/.rtbz files flat in that directory, and make sure the chesstrainer
user can read them:
sudo install -d -o chesstrainer -g chesstrainer /opt/chesstrainer/tablebases
sudo chown -R chesstrainer:chesstrainer /opt/chesstrainer/tablebases
sudo chmod -R u=rwX,g=rX,o= /opt/chesstrainer/tablebases
Set the path in /etc/chesstrainer/env:
TABLEBASE_PATH=/opt/chesstrainer/tablebases
Then restart the service, because TABLEBASE_PATH is read at boot:
sudo systemctl restart chesstrainer
Check /health. It should move from degraded / not_configured to
ok / local:
curl -s http://127.0.0.1:4100/health
sudo journalctl -u chesstrainer -n 50 --no-pager | grep -i tablebase
If the path is missing, unreadable, or empty, the app falls back to Lichess and
health stays degraded but still returns 200.
To refresh the files:
mix tablebase.fetch --path /opt/chesstrainer/tablebases
sudo chown -R chesstrainer:chesstrainer /opt/chesstrainer/tablebases
sudo systemctl restart chesstrainer
First-time setup
Run as root or with sudo:
useradd --system --user-group \
--home-dir /opt/chesstrainer \
--create-home \
--shell /usr/sbin/nologin \
chesstrainer
install -d -o chesstrainer -g chesstrainer /opt/chesstrainer/releases
install -d -o chesstrainer -g chesstrainer /opt/chesstrainer/src
install -d -m 755 /etc/chesstrainer
install -m 600 /dev/null /etc/chesstrainer/env
install -m 600 /dev/null /etc/chesstrainer/postgres.env
chown root:root /etc/chesstrainer/env /etc/chesstrainer/postgres.env
Edit both files with sudoedit and set the generated secrets and
DATABASE_URL.
Unified Compose: add Postgres
Add the chesstrainer-db service and chesstrainer_pgdata volume from
deploy/postgres.compose.yml to the VPS's existing unified
docker-compose.yml. Do not run the reference file as a separate project.
The image is pinned directly in the unified compose:
image: postgres:18.6
For an immutable pin, replace it with the exact digest:
image: postgres:18.6@sha256:...
There is no POSTGRES_IMAGE environment variable involved.
Because postgres.env is root-owned 0600, run Compose with sudo:
sudo docker compose -f <UNIFIED_COMPOSE> config >/dev/null
sudo docker compose -f <UNIFIED_COMPOSE> up -d chesstrainer-db
sudo docker compose -f <UNIFIED_COMPOSE> ps chesstrainer-db
sudo docker compose -f <UNIFIED_COMPOSE> logs --tail=50 chesstrainer-db
Verify the database:
sudo docker compose -f <UNIFIED_COMPOSE> exec -T chesstrainer-db \
psql -U chesstrainer -d chesstrainer_prod -c 'select 1;'
To change the Postgres version:
# edit the image: line in the unified compose
sudo docker compose -f <UNIFIED_COMPOSE> pull chesstrainer-db
sudo docker compose -f <UNIFIED_COMPOSE> up -d chesstrainer-db
Read-only database access
The app connects as chesstrainer, which is a superuser. Use a separate
read-only role for inspection, debugging, and dashboards.
Create it once per environment:
sudo docker compose -f <UNIFIED_COMPOSE> exec -T chesstrainer-db \
psql -U chesstrainer -d chesstrainer_prod <<'SQL'
CREATE ROLE chesstrainer_ro LOGIN PASSWORD '<generated-password>';
GRANT CONNECT ON DATABASE chesstrainer_prod TO chesstrainer_ro;
GRANT USAGE ON SCHEMA public TO chesstrainer_ro;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO chesstrainer_ro;
-- Cover tables created by future migrations. Migrations run as chesstrainer.
ALTER DEFAULT PRIVILEGES FOR ROLE chesstrainer IN SCHEMA public
GRANT SELECT ON TABLES TO chesstrainer_ro;
-- Optional: read other sessions and server settings in pg_stat_* views.
GRANT pg_read_all_stats TO chesstrainer_ro;
SQL
Generate the password with openssl rand -hex 32 and keep it in a password
manager. The only access it has is SELECT; there are no INSERT, UPDATE,
DELETE, or object-creation grants. If the role already exists, skip the
CREATE ROLE line and re-run the grants.
Optional defence in depth:
ALTER ROLE chesstrainer_ro SET default_transaction_read_only = on;
This is not a hard boundary (the role can override it per session), but it turns an accidental write into an error.
Connect from the VPS
docker exec connects over the container's Unix socket, which the Postgres
image trusts locally, so no password is needed:
sudo docker exec -it chesstrainer-db \
psql -U chesstrainer_ro -d chesstrainer_prod
Connect from a workstation over SSH
Postgres is published on 127.0.0.1:5432 only. OpenSSH also disables
forwarding by default in /etc/ssh/sshd_config (AllowTcpForwarding no).
Allow local forwarding for the deploy user only, restricted to Postgres, by
appending this to the end of that file:
Match User <username>
AllowTcpForwarding local
PermitOpen 127.0.0.1:5432
Then validate and reload:
sudo sshd -t
sudo systemctl reload ssh
sudo sshd -T -C user=<username>,host=localhost,addr=127.0.0.1 | grep -iE 'allowtcpforwarding|permitopen'
Open the tunnel (local port 5433 because the dev database already binds
5432):
ssh -N -L 5433:127.0.0.1:5432 <VPS_USER>@<VPS_HOST>
Connect from a second terminal:
psql "postgres://chesstrainer_ro@127.0.0.1:5433/chesstrainer_prod"
Install postgresql-client if psql is not available on the workstation.
To avoid the password prompt, add this to ~/.pgpass (mode 0600):
127.0.0.1:5433:chesstrainer_prod:chesstrainer_ro:<password>
DBeaver uses the same tunnel. Create a PostgreSQL connection with host
127.0.0.1, port 5432, database chesstrainer_prod, user
chesstrainer_ro, then on the SSH tab enable the tunnel with host
<VPS_HOST>, port 22, user <VPS_USER>, and a public key. The
AllowTcpForwarding local rule above is required for any tunnelled client.
Reset the role password:
sudo docker compose -f <UNIFIED_COMPOSE> exec -T chesstrainer-db \
psql -U chesstrainer -d chesstrainer_prod \
-c "ALTER ROLE chesstrainer_ro PASSWORD '<new-password>';"
Verify the role is read-only:
psql "postgres://chesstrainer_ro@127.0.0.1:5433/chesstrainer_prod" \
-c 'select current_user, current_database();'
psql "postgres://chesstrainer_ro@127.0.0.1:5433/chesstrainer_prod" \
-c 'create table should_fail (i int);'
The second command must fail with permission denied for schema public.
systemd unit
From the deploy artifact copy:
sudo install -o root -g root -m 644 /opt/chesstrainer/src/deploy/chesstrainer.service \
/etc/systemd/system/chesstrainer.service
sudo systemctl daemon-reload
sudo systemctl enable chesstrainer
The unit includes ConditionPathExists=/opt/chesstrainer/current/bin/server, so
it is safe to enable before the first release. systemd skips it until a release
is in place.
The backup unit and timer are installed separately; see §Backups.
Caddy
Add the site block from deploy/Caddyfile by hand. It imports
cloudflare-ranges.caddy, which is generated by
deploy/update-cloudflare-ranges.sh and refreshed by a systemd timer.
Caddy must only accept Cloudflare traffic and must replace X-Forwarded-For
with Cloudflare's CF-Connecting-IP value:
innersquare.app {
encode zstd gzip
@cloudflare {
import cloudflare-ranges.caddy
}
handle @cloudflare {
reverse_proxy 127.0.0.1:4100 {
header_up X-Forwarded-For {http.request.header.CF-Connecting-IP}
}
}
handle {
respond "Access Denied" 403
}
}
www.innersquare.app {
redir https://innersquare.app{uri} permanent
}
Install the range file and the hourly refresh timer:
sudo install -m 0755 /opt/chesstrainer/src/deploy/update-cloudflare-ranges.sh \
/usr/local/sbin/chesstrainer-update-cloudflare-ranges
sudo install -m 0644 /opt/chesstrainer/src/deploy/cloudflare-ranges.caddy \
/etc/caddy/cloudflare-ranges.caddy
sudo install -m 0644 /opt/chesstrainer/src/deploy/chesstrainer-cloudflare-ranges.service \
/etc/systemd/system/chesstrainer-cloudflare-ranges.service
sudo install -m 0644 /opt/chesstrainer/src/deploy/chesstrainer-cloudflare-ranges.timer \
/etc/systemd/system/chesstrainer-cloudflare-ranges.timer
The installed pieces are:
| Path | Mode | Purpose |
|---|---|---|
/usr/local/sbin/chesstrainer-update-cloudflare-ranges |
0755 |
fetch/validate/reload script |
/etc/caddy/cloudflare-ranges.caddy |
0644 |
generated remote_ip list |
/etc/systemd/system/chesstrainer-cloudflare-ranges.service |
0644 |
oneshot service |
/etc/systemd/system/chesstrainer-cloudflare-ranges.timer |
0644 |
hourly timer |
If /opt/chesstrainer/src/deploy is not present on the VPS, create the same
four files directly with sudo tee from the canonical contents in deploy/.
The timer only depends on the installed copies; it does not read the repo.
Run it once manually and enable the timer:
sudo systemctl daemon-reload
sudo systemctl enable --now chesstrainer-cloudflare-ranges.timer
sudo systemctl start chesstrainer-cloudflare-ranges.service
sudo journalctl -u chesstrainer-cloudflare-ranges.service -n 50 --no-pager
The script fetches https://www.cloudflare.com/ips-v4 and https://www.cloudflare.com/ips-v6, validates the result, and reloads Caddy only when the ranges actually changed. If the download, Caddy validation or reload fails, it keeps or restores the previous file.
Check the timer and last run:
systemctl list-timers chesstrainer-cloudflare-ranges.timer
systemctl status chesstrainer-cloudflare-ranges.service
journalctl -u chesstrainer-cloudflare-ranges.service --since "24 hours ago"
Do not edit /etc/caddy/cloudflare-ranges.caddy by hand; the next timer run
overwrites it.
To update the refresher itself, replace the installed script or unit files from the deploy directory, then:
sudo systemctl daemon-reload
sudo systemctl restart chesstrainer-cloudflare-ranges.timer
sudo systemctl start chesstrainer-cloudflare-ranges.service
Without this, TRUST_FORWARDED_FOR=true records a Cloudflare edge address
instead of the visitor, and the login throttle and guest puzzle limit bucket
whole Cloudflare POPs together.
If a new VPS uses imports, put the site block in its own file instead; keep the
import cloudflare-ranges.caddy path next to the Caddyfile.
Do not switch Caddy until a release is running and verified on
127.0.0.1:4100. Before that, keep the static coming-soon site live.
Validate and reload:
sudo caddy validate --config /etc/caddy/Caddyfile
sudo systemctl reload caddy
A direct request to the origin with a forged CF-Connecting-IP header should
receive 403 Access Denied.
Do not let a deploy overwrite the shared Caddyfile.
See docs/security.md for the application throttle, real-IP setup and the
Cloudflare edge controls. Production currently has Bot Fight Mode off (see
#153) and the free-plan login flood rule: 10 POSTs per 10 seconds per IP,
blocked for 10 seconds.
Building a release
Build on a machine matching Ubuntu 24.04 / glibc 2.39, or let CI do it. The release tarball is the contents of the release directory:
npm ci --prefix assets --no-audit --no-fund
export MIX_ENV=prod
mix deps.get --only prod
mix assets.deploy
mix release
tar -czf chesstrainer-release.tar.gz -C _build/prod/rel/chesstrainer .
npm ci is required before mix assets.deploy: Tailwind imports the
Chessground CSS and esbuild bundles the package from assets/node_modules.
mix assets.deploy runs compile first so the generated LiveView colocated
CSS exists.
First deploy and Caddy switch
Upload the tarball:
scp chesstrainer-release.tar.gz <VPS_USER>@<VPS_HOST>:/tmp/
On the VPS:
sudo /opt/chesstrainer/src/deploy/activate /tmp/chesstrainer-release.tar.gz
The script:
- Extracts into
/opt/chesstrainer/releases/<timestamp> - Fixes ownership
- Takes a verified pre-deploy database dump, aborting if it fails (see §Backups)
- Flips
/opt/chesstrainer/current - Restarts
chesstrainer.service - Migrations run via
ExecStartPre - Polls
/health - Rolls back to the previous release on failure
- Prunes old releases, keeping 5 by default
Verify locally before touching Caddy:
curl -i http://127.0.0.1:4100/health
curl -i http://127.0.0.1:4100/
sudo journalctl -u chesstrainer -n 100 --no-pager
Then add the Caddy block and reload. Verify publicly:
curl -I https://innersquare.app/
curl -I https://innersquare.app/privacy
curl -I https://innersquare.app/terms-conditions
Subsequent deploys
scp chesstrainer-release.tar.gz <VPS_USER>@<VPS_HOST>:/tmp/
ssh <VPS_USER>@<VPS_HOST>
sudo /opt/chesstrainer/src/deploy/activate /tmp/chesstrainer-release.tar.gz
Optional overrides:
RELEASE_ID=20261109120000 \
HEALTH_URL=http://127.0.0.1:4100/health \
KEEP_RELEASES=5 \
sudo -E /opt/chesstrainer/src/deploy/activate /tmp/chesstrainer-release.tar.gz
Every activation takes a verified pre-deploy dump before migrations run. If
Postgres is intentionally unavailable and you accept the risk, add
SKIP_PRE_DEPLOY_BACKUP=1 to the override list; see §Backups.
GitHub Actions CD
Merging to main (or running the workflow manually) tests, builds, uploads,
and deploys the release. The deploy job uses SSH as <username>.
Deploy key
Generate a dedicated key, not your personal one:
ssh-keygen -t ed25519 -C "github-actions-chesstrainer" \
-f ~/.ssh/chesstrainer_deploy -N ""
ssh-copy-id -i ~/.ssh/chesstrainer_deploy.pub <username>@<VPS_HOST>
VPS_HOST must be the public IP or DNS name reachable from GitHub. Do not use
a local SSH alias such as personal-projects.
Passwordless sudo for the deploy scripts
sudo visudo -f /etc/sudoers.d/chesstrainer-deploy
<username> ALL=(root) NOPASSWD: /opt/chesstrainer/src/deploy/activate, /opt/chesstrainer/src/deploy/rollback
Validate:
sudo chown root:root /etc/sudoers.d/chesstrainer-deploy
sudo chmod 0440 /etc/sudoers.d/chesstrainer-deploy
sudo visudo -c
sudo -n -l /opt/chesstrainer/src/deploy/activate
sudo -n -l /opt/chesstrainer/src/deploy/rollback
GitHub environment
Create the production environment, optionally with required reviewers, and add environment secrets:
| Secret | Value |
|---|---|
VPS_HOST |
public VPS IP or DNS name |
VPS_USER |
<username> |
VPS_SSH_KEY |
private deploy key |
VPS_KNOWN_HOSTS |
ssh-keyscan -H <VPS_HOST> output |
Generate known_hosts with:
ssh-keyscan -H <VPS_HOST>
Workflow
The workflow:
- Runs
mix precommitagainst Postgres. - Builds the release artifact on
ubuntu-24.04. - On
mainorworkflow_dispatch, uploads the tarball, installs the SSH key, runssudo -n .../activate, and performs public smoke checks. Activation takes a pre-deploy dump first and fails if it cannot. - A failed activation health check rolls back to the previous release.
Manual dispatch is available from the Actions UI or:
gh workflow run ci.yml --ref main
Rollback
sudo /opt/chesstrainer/src/deploy/rollback
Or target a specific release:
sudo /opt/chesstrainer/src/deploy/rollback 20261109120000
The script flips current, restarts the service, and checks /health.
Rollback reverts the release, not the database: migrations are not reversed. If a migration changed data, restore the matching pre-deploy dump instead; see §Backups.
Logs and health
sudo systemctl status chesstrainer
sudo journalctl -u chesstrainer -n 100 --no-pager
curl -i http://127.0.0.1:4100/health
/health returns JSON with status, db, tablebase, and process.
It returns 200 for ok/degraded and 503 for down. A missing local tablebase
reports degraded but still 200 because Lichess fallback is available. See
§Monitoring and alerting for the public probe that watches it, and for what a
green check does and does not prove.
Reading production logs
Production emits one JSON object per line. Use -o cat so journald prints the
raw line instead of its own field wrapper, then pipe to jq:
# Recent logs, pretty-printed
sudo journalctl -u chesstrainer -o cat -n 50 --no-pager | jq .
# Errors only
sudo journalctl -u chesstrainer -o cat --since "1 hour ago" | jq 'select(.level == "error")'
# Everything from one request
sudo journalctl -u chesstrainer -o cat | jq 'select(.request_id == "REQUEST_ID")'
# Crash reasons and stacktraces
sudo journalctl -u chesstrainer -o cat | jq -r 'select(.error) | .error.reason, (.error.stacktrace[]?)'
Each line has timestamp, level and message, plus request_id when one
exists. Access lines add method, path, status and duration_ms; crashes
add an error object with kind, reason and stacktrace. Query parameters
and request bodies are never logged, token-shaped path segments are redacted,
and email addresses or credentials in metadata render as [REDACTED].
The full field table, redaction rules and examples for adding background logs
are in docs/logging.md.
Dev and test still use the readable [level] message formatter.
Keep the journal
By default some distributions keep journald logs in memory only. On the VPS, make sure they survive a reboot and have a retention budget:
sudo mkdir -p /var/log/journal
sudo systemctl restart systemd-journald
journalctl --disk-usage
Set Storage=persistent and a SystemMaxUse= value in
/etc/systemd/journald.conf if the distribution does not already. journald
rate-limits bursts; if a crash storm drops lines, raise RateLimitBurst or
RateLimitIntervalSec there.
Monitoring and alerting
Three checks watch production at three different layers:
| Where | What | Notices | Alerts |
|---|---|---|---|
| VPS | chesstrainer-probe.timer → chesstrainer-probe.service |
the user-facing HTTP path: DNS, Cloudflare, Caddy, Phoenix, Postgres | healthchecks.io |
| VPS | Beszel hub + agent | host liveness, CPU, memory, disk, container restarts and OOM kills | Beszel notifications |
| VPS | chesstrainer-backup.timer heartbeat |
whether the nightly dump finished | healthchecks.io (its own check) |
The backup heartbeat is documented in §Backups. The probe and the Beszel gap are below.
On-box prober, off-box alerting
The probe runs on the VPS and curls the public hostname, so it crosses the same path a visitor does: Cloudflare's DNS and edge, the proxy back to Caddy, Caddy's Cloudflare allowlist, the Phoenix router, and one database query.
That is deliberately not 127.0.0.1:4100. A loopback probe stays green through
every failure that makes the site unusable: a bad cloudflare-ranges.caddy
reload, a stopped Caddy (521), an expired origin certificate (526), a
deleted proxied DNS record, a WAF rule blocking /health, or a Cloudflare
incident. Beszel is green through all of those too, because the host and the
container are fine.
The trade-off is that the prober shares a host with the app. It cannot report that its own host is down, and it cannot prove the view from another network. Host death is caught by silence: the timer stops, no ping arrives, and healthchecks.io — which is off-box — alerts after the grace period. That is the whole reason the check-in URL points at a SaaS rather than at another box here. Do not describe this as external uptime monitoring.
The healthchecks.io check
Create the check at healthchecks.io:
- Name
innersquare.app /health, period 2 minutes, grace 2 minutes - Integrations: email and Discord (webhook). Use each integration's test notification once, and keep recovery notifications enabled so a resolved incident is visible.
- Copy the ping URL. It is a secret: anyone who has it can mark the check green. It belongs only in the root-owned env file.
Every check-in carries evidence:
- A healthy run posts the raw
/healthJSON. - A tablebase mismatch still posts a success, because a Lichess fallback is not an outage. The body carries the expected and actual values, so the mismatch is visible in the check history without alerting anyone.
- A non-200, a malformed body, a
dbthat is notok, or a network/TLS failure posts to<ping-url>/failwith a JSON body containing the reason, the HTTP status, the curl exit code, the response body, and a loopbackorigin_status/origin_bodyfor comparison. The origin fields are evidence only; the verdict comes from the public request.
A ping that cannot be delivered is a journal warning, not a second verdict: the next timer run reports the state again.
Install the probe
Write the env file, root-owned and 0600:
sudo install -m 600 /dev/null /etc/chesstrainer/probe.env
sudoedit /etc/chesstrainer/probe.env
Start from deploy/probe.env.example. PROBE_URL and HEALTHCHECK_URL are the
only required values. Then install the script, the units, and the timer:
sudo install -m 0755 /opt/chesstrainer/src/deploy/probe \
/usr/local/sbin/chesstrainer-probe
sudo install -m 0644 \
/opt/chesstrainer/src/deploy/chesstrainer-probe.service \
/opt/chesstrainer/src/deploy/chesstrainer-probe.timer \
/etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now chesstrainer-probe.timer
The timer runs the installed /usr/local/sbin/chesstrainer-probe, not the copy
under /opt/chesstrainer/src, because that tree belongs to the application
user. When deploy/probe changes, install it to both places so the timer and
the repo agree.
Verify:
systemctl list-timers chesstrainer-probe.timer
sudo journalctl -u chesstrainer-probe -n 5 -o cat
Expect a probe.ok line every two minutes and a green check within one period.
Each run prints exactly one JSON line (probe.ok, probe.degraded or
probe.failed), like the backup script, so #140's log shipping can pick it up.
What /health does and does not mean
/health returns status, db, tablebase and process. It returns 200 for
ok and degraded, and 503 for down.
dbis oneSELECT 1through the connection pool. It proves the pool can reach Postgres. It does not prove migrations are current or that writes work.tablebasereflects the local Syzygy directory.fallbackmeans lookups go to Lichess.degradedis 200 because the fallback is usable.processis always"ok". It is not independent evidence of liveness.
A green probe therefore means: DNS resolved, Cloudflare served the request, Caddy routed it under the Cloudflare allowlist, Phoenix answered, and Postgres answered one query. It does not mean background jobs are running, SES accepts mail, writes succeed, tablebase lookups are fast, or the site is fast under load. Beszel covers resource exhaustion and container restarts; the release health check and the logs cover deploy mistakes.
Beszel's blind spot
Beszel watches the host: agent connection, CPU, memory, disk, container
restarts, OOM kills. Its hub runs on the same VPS as the agent, so a VPS
outage or network loss silences it as well — that case is covered by the probe
check going quiet. More importantly, Beszel never calls /health. A Phoenix
release returning 500s on a healthy host, or Postgres refusing connections
inside a running container, reads green in Beszel and red in the probe. The two
are complementary, not redundant.
Drills
Nothing below touches production data, and the app keeps serving throughout.
--dry-run needs no network and no env file. It evaluates the success,
degraded, fail, non-JSON and unreachable fixtures and sends no check-ins:
sudo /usr/local/sbin/chesstrainer-probe --dry-run
A degraded check-in, recorded but never alerting. A value passed on the command line wins over the env file:
sudo EXPECT_TABLEBASE=not_configured /usr/local/sbin/chesstrainer-probe
Expect exit 0, a green check, and a body showing tablebase: local against
expected_tablebase: not_configured. No email and no Discord message.
The one fail ping, and the full alert path. Stop the timer first so the next scheduled run cannot mask the result:
sudo systemctl stop chesstrainer-probe.timer
sudo PROBE_URL=https://innersquare.app/health-does-not-exist \
/usr/local/sbin/chesstrainer-probe
Expect exit 1, one probe.failed line with http_status: 404, one fail ping
whose body shows origin_status: 200, one alert in Discord/email, and no
second fail ping. Then send a real check-in and re-arm:
sudo /usr/local/sbin/chesstrainer-probe
sudo systemctl start chesstrainer-probe.timer
The real run posts a success, and healthchecks.io notifies that the check is
back up. For the unreachable branch, point the probe at a closed port instead;
the origin evidence should still show 200:
sudo PROBE_URL=https://127.0.0.1:9/health /usr/local/sbin/chesstrainer-probe
The outage drill: stop Phoenix, prove the alert, bring it back. Forcing each probe run instead of waiting for the timer keeps the outage under a minute. Start the probe once per step: every run sends its own check-in, so a second start during the outage is a second alert.
sudo systemctl start chesstrainer-probe.service # baseline: probe.ok
sudo systemctl stop chesstrainer
sudo systemctl start chesstrainer-probe.service || true # probe.failed, http 502
# confirm one Discord message and one email, then:
sudo systemctl start chesstrainer
curl -sS -o /dev/null -w '%{http_code}\n' https://innersquare.app/health # 200
sudo systemctl start chesstrainer-probe.service # probe.ok, recovery
Expect http_status: 502 from Caddy and origin_status: 000 with
origin_curl_exit: 7 in the fail body: Phoenix is down, while the host and the
edge are fine.
Planned maintenance: stopping the timer is reported as a down check after period plus grace, so about four minutes. Pause the check in the healthchecks.io UI for the duration, or accept the alert.
The probe reaches the edge as a plain curl. If a bot policy or WAF rule ever
starts rejecting it, the fail body's origin_status shows the app itself is
healthy; that is an edge-policy change, not an application outage. Re-enabling
Bot Fight Mode is tracked in #153.
Drill log
- 2026-11-13 — install and first drills. Script and units installed 12:34
UTC, timer enabled, first scheduled
probe.okat 12:35. Forced-failure drill (PROBE_URLat a missing path): one fail ping withhttp_status: 404andorigin_status: 200, one Discord message and one email, then one recovery notification. Degraded drill (EXPECT_TABLEBASE=not_configured): green check-in with the mismatch in the body, no notification. Outage drill: the app was stopped 13:32:55 and started 13:33:42 UTC; the manual probe was started twice while it was down, soprobe.failedfired at 13:33:06 and 13:33:14 (http 502) and two notifications arrived for one down → up transition.probe.okat 13:34:03 and one recovery notification. No production data was touched. Start the probe once per drill step.
Interactive shell and user management
Attach an IEx shell to the running release. This does not restart the service and does not require direct database access:
sudo /opt/chesstrainer/current/bin/chesstrainer remote
You should get a prompt like iex(chesstrainer@<host>)1>. Exit with Ctrl-C
twice. Never call System.halt/0, :init.stop/0, or
Application.stop(:chesstrainer) from this shell — those stop production.
Use it for account and role management. Go through the application contexts (not SQL) so changesets, bcrypt hashing, timestamps, and token bookkeeping stay consistent:
# Look up by email or username
u = Chesstrainer.Accounts.get_user_by_login("owen")
# Promote or demote
{:ok, u} = Chesstrainer.Accounts.update_user_role(u, :super_admin)
# Confirm an account without sending email
{:ok, u} = Chesstrainer.Repo.update(Chesstrainer.Accounts.User.confirm_changeset(u))
# Create an account, generating its password server-side
pw = :crypto.strong_rand_bytes(18) |> Base.url_encode64(padding: false)
{:ok, u} =
Chesstrainer.Accounts.register_user(%{
email: "you@example.com",
username: "owen",
password: pw
})
{:ok, u} = Chesstrainer.Repo.update(Chesstrainer.Accounts.User.confirm_changeset(u))
pw
Notes:
register_user/1hashes the password with bcrypt;passwordis virtual and is never written to the database.- Do not type passwords into the shell if you can avoid it — IEx may keep a
history file — and never paste a password from shell history. Generate one as
above and change it later at
/users/profile/settings. mix accounts.promotestill works when a compiled source checkout is present on the host, but the release shell is the supported production path.- There is no guard against demoting the last
super_adminoutside the future admin UI, so check the result before demoting.
If remote fails while booting its client VM (for example
Kernel pid terminated (logger) with code_server/persistent_term in the
trace), the deployed release is stale, mixed, or missing IEx: rebuild and
redeploy it. To attach in the meantime, use the release's own ERTS:
sudo bash -c '
COOKIE=$(cat /opt/chesstrainer/current/releases/COOKIE)
exec /opt/chesstrainer/current/erts-*/bin/erl \
-sname debug -setcookie "$COOKIE" -remsh chesstrainer
'
That shell speaks Erlang syntax, so quote Elixir module atoms, e.g.
'Elixir.Chesstrainer.Accounts':register_user(...).
Backups
Backups are produced on the VPS, encrypted there, and pulled to a workstation. Nothing leaves the VPS unencrypted, and the VPS never holds the key that could read the archives.
What runs
| Where | What | When | Retention |
|---|---|---|---|
| VPS | chesstrainer-backup.timer → chesstrainer-backup.service |
daily 03:15 UTC, Persistent=true, 5 min jitter |
3 days plaintext, 14 days encrypted |
| VPS | the same script, called by deploy/activate |
before every migration | last 5 pre-deploy sets |
| Workstation | chesstrainer-backup-pull.timer |
daily 12:00 local, 10 min after boot | 7 daily, 4 weekly, 6 monthly |
| Workstation | chesstrainer-backup-drill.timer |
monthly on the 1st at 09:00 | drills.log is never pruned |
Each run writes one JSON line (backup.completed, backup.failed,
backup.skipped, pull.completed, pull.failed, pull.skipped,
drill.completed, drill.failed) so #140's log shipping can alert on it. A
skipped line means another run held the lock; it is not an alarm.
What a dump contains
<ts>.dump—pg_dump -Fcofchesstrainer_prod<ts>.globals.sql—pg_dumpall --globals-only, enough to recreatechesstrainerandchesstrainer_roon a fresh cluster<ts>.counts— row counts forusers,endgames,attempts, used by the drill<ts>.sha256— checksums for the three files above
Not backed up, by design: tablebases and the tablebase cache (both are
re-fetchable), anything under priv/static, and release tarballs. Dumps contain
user emails and bcrypt hashes: they are 0600 root while plaintext, encrypted
before they leave the box, and pruned aggressively.
RPO and RTO
RPO is 24 hours on the VPS and "the age of the newest pulled archive" off-box. The pull runs daily, so a laptop that is off for a day leaves the newest off-site copy around 48 hours old; the VPS keeps 14 days of encrypted archives, which bounds how far behind the pull can fall. RTO is under an hour: create the database, restore the newest dump, activate a release (see Disaster recovery).
One-time setup: VPS
Install age. Encryption happens on the server; only the workstation can
decrypt.
sudo apt install age
Create the configuration file. Generate AGE_RECIPIENT on the workstation
first (next section) and paste the public key.
sudo install -m 600 /dev/null /etc/chesstrainer/backup.env
sudoedit /etc/chesstrainer/backup.env
Start from deploy/backup.env.example. The required values are AGE_RECIPIENT
and HEALTHCHECK_URL.
Create a healthchecks.io check with a period of 1 day and a grace time of 4
hours, then paste its ping URL into HEALTHCHECK_URL. The script pings it after
a good run and <url>/fail after a bad one, so silence means "no backup
finished", even when the VPS itself is down.
Install the script and units:
sudo install -m 0755 /opt/chesstrainer/src/deploy/backup \
/usr/local/sbin/chesstrainer-backup
sudo install -m 0644 \
/opt/chesstrainer/src/deploy/chesstrainer-backup.service \
/opt/chesstrainer/src/deploy/chesstrainer-backup.timer \
/opt/chesstrainer/src/deploy/chesstrainer-backup-failure.service \
/etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable --now chesstrainer-backup.timer
The timer runs the installed /usr/local/sbin/chesstrainer-backup, not the
copy under /opt/chesstrainer/src, because that tree is writable by the
application user. When deploy/backup changes, copy it to both places so the
timer and deploy/activate use the same script.
Run it once and inspect the result:
sudo systemctl start chesstrainer-backup
sudo journalctl -u chesstrainer-backup -n 20 --no-pager -o cat
sudo ls -l /var/backups/chesstrainer /var/backups/chesstrainer/offsite
systemctl list-timers chesstrainer-backup.timer
Expect a .dump, .globals.sql, .counts and .sha256 at mode 0600, and
one .tar.age in offsite/ at mode 0644. offsite/ is the only thing the
workstation reads, which is why the pull needs no sudo. The vault directory is
0711, so the workstation user can traverse into offsite/ but cannot list the
plaintext dumps.
One-time setup: workstation
sudo apt install age rsync
mkdir -p ~/.config/chesstrainer
chmod 700 ~/.config/chesstrainer
age-keygen -o ~/.config/chesstrainer/age.key
chmod 600 ~/.config/chesstrainer/age.key
grep 'public key' ~/.config/chesstrainer/age.key
Copy that public key into AGE_RECIPIENT on the VPS. Back up the private
key (password manager, printed copy, offline disk): without it every archive
is unreadable, and the VPS deliberately cannot decrypt them.
Create the pull configuration:
install -m 600 /dev/null ~/.config/chesstrainer/backup-pull.env
$EDITOR ~/.config/chesstrainer/backup-pull.env
Start from deploy/workstation/backup-pull.env.example. DEST should point at
a real disk, for example /mnt/Mint Data/chesstrainer-backups. The script
refuses to run when DEST's parent is missing, so an unmounted disk cannot
silently fill the root filesystem.
Set PULL_HEALTHCHECK_URL to a second healthchecks.io check with a long period
(for example 72 hours) if you want to hear about it when the workstation stops
pulling. Leave it empty for journal-only reporting.
Install the scripts and user units:
REPO=~/Desktop/chess-trainer # wherever this repo is cloned
install -d ~/.local/bin
install -m 0755 "$REPO/deploy/workstation/backup-pull" ~/.local/bin/chesstrainer-backup-pull
install -m 0755 "$REPO/deploy/workstation/backup-drill" ~/.local/bin/chesstrainer-backup-drill
install -d ~/.config/systemd/user
install -m 0644 \
"$REPO/deploy/workstation/chesstrainer-backup-pull.service" \
"$REPO/deploy/workstation/chesstrainer-backup-pull.timer" \
"$REPO/deploy/workstation/chesstrainer-backup-drill.service" \
"$REPO/deploy/workstation/chesstrainer-backup-drill.timer" \
"$REPO/deploy/workstation/chesstrainer-backup-alert@.service" \
~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now chesstrainer-backup-pull.timer
systemctl --user enable --now chesstrainer-backup-drill.timer
sudo loginctl enable-linger "$USER"
enable-linger lets the user timers run while the laptop is on but you are not
logged in. The pull also fires ten minutes after boot, and Persistent=true
catches up after the machine was off. Verify lingering stuck:
loginctl show-user "$USER" -p Linger # Linger=yes
Run the first pull and drill by hand:
~/.local/bin/chesstrainer-backup-pull
ls -l "/mnt/Mint Data/chesstrainer-backups"
~/.local/bin/chesstrainer-backup-drill
cat "/mnt/Mint Data/chesstrainer-backups/drills.log"
A passing drill line looks like:
{"timestamp":"2026-11-11T21:05:46Z","event":"drill.completed","archive":"20261111T120000Z.tar.age","result":"pass","counts":"users=3,endgames=42,attempts=900","duration_s":3}
The drill creates a throwaway postgres:18.6 container, applies the globals
dump, asserts that chesstrainer and chesstrainer_ro exist, restores into
chesstrainer_restore, and compares the three row counts with the .counts
file from dump time. Restored counts may be higher (the counts are taken moments
before the dump); lower counts fail the drill.
Retention
| Location | Plaintext | Encrypted | Notes |
|---|---|---|---|
VPS /var/backups/chesstrainer |
3 days | 14 days | pre-deploy sets: last 5 |
Workstation DEST |
never | 7 daily, 4 weekly, 6 monthly | monthly copies outlive the VPS copies |
The VPS is a staging area for the workstation; weekly and monthly generations accumulate there. If the laptop is off for more than 14 days, the archives the VPS dropped in the meantime are gone — that is the documented off-site RPO trade-off. Moving to an always-on destination (a second host or object storage with restic) is tracked in #151.
Restore a dump
Always restore into a scratch database first. From an archive on the workstation:
WORK=$(mktemp -d)
age --decrypt -i ~/.config/chesstrainer/age.key \
-o "$WORK/dump.tar" \
"/mnt/Mint Data/chesstrainer-backups/<timestamp>.tar.age"
tar -xf "$WORK/dump.tar" -C "$WORK"
(cd "$WORK" && sha256sum --check --strict <timestamp>.sha256)
Then restore into a scratch container:
docker run --rm -d --name restore-check \
-e POSTGRES_PASSWORD=restore -e POSTGRES_HOST_AUTH_METHOD=trust postgres:18.6
docker exec restore-check psql -U postgres -c 'create database chesstrainer_restore'
docker exec -i restore-check pg_restore -U postgres -d chesstrainer_restore \
--single-transaction --exit-on-error < "$WORK/<timestamp>.dump"
docker exec restore-check psql -U postgres -d chesstrainer_restore \
-c 'select count(*) from users; select count(*) from attempts;'
docker rm -f restore-check
rm -rf "$WORK"
Disaster recovery
- Build a new VPS following this README through First-time setup, without activating a release or switching Caddy.
- Decrypt and unpack the newest archive on the workstation (above).
- Copy the dump and globals to the new VPS:
scp "$WORK/<timestamp>.dump" "$WORK/<timestamp>.globals.sql" <VPS_USER>@<VPS_HOST>:/tmp/ - Recreate the cluster roles. Connect as
chesstrainer: it is the container's superuser (POSTGRES_USER), and the dump's final grants are issuedGRANTED BY chesstrainer, which another role cannot replay. The line that recreateschesstraineritself reports "already exists" and is harmless:sudo docker exec -i chesstrainer-db psql -U chesstrainer < /tmp/<timestamp>.globals.sql - Create the database and restore it:
sudo docker exec chesstrainer-db createdb -U chesstrainer chesstrainer_prodsudo docker exec -i chesstrainer-db pg_restore -U chesstrainer -d chesstrainer_prod --single-transaction --exit-on-error < /tmp/<timestamp>.dump - Verify the row counts, activate a release, then restore the Caddy block.
- Take a fresh backup (
sudo systemctl start chesstrainer-backup) and pull so the new VPS is covered immediately.
pg_dump and pg_restore must come from the container's Postgres 18, not the
host.
Pre-deploy dump
deploy/activate calls chesstrainer-backup with BACKUP_TAG=predeploy after
extracting the release and before flipping current, so the state before
bin/migrate runs in ExecStartPre is always recoverable. A failed dump aborts
the activation. The escape hatch is deliberate and should be rare:
sudo SKIP_PRE_DEPLOY_BACKUP=1 /opt/chesstrainer/src/deploy/activate /tmp/chesstrainer-release.tar.gz
If the installed script is missing, activate uses the backup next to it; if
neither exists it refuses to activate rather than migrate without a dump. The
GitHub Actions deploy job inherits the same behaviour.
Run it by hand
Every timer can be triggered directly. On the VPS:
# exactly what the timer runs
sudo systemctl start chesstrainer-backup
sudo journalctl -u chesstrainer-backup -n 20 -o cat
# an ad-hoc dump with its own tag
sudo BACKUP_TAG=manual /usr/local/sbin/chesstrainer-backup
# schedule, last result and last exit
systemctl list-timers chesstrainer-backup.timer
systemctl show chesstrainer-backup.service -p Result -p ExecMainExitTimestamp
On the workstation:
# pull now, or through the unit (same PATH and OnFailure handling)
~/.local/bin/chesstrainer-backup-pull
systemctl --user start chesstrainer-backup-pull.service
# drill the newest archive, or a named one
~/.local/bin/chesstrainer-backup-drill
~/.local/bin/chesstrainer-backup-drill 20261111T213734Z.tar.age
# results and schedule
cat "/mnt/Mint Data/chesstrainer-backups/drills.log"
systemctl --user list-timers 'chesstrainer-backup-*'
A backup.skipped or pull.skipped line means another run held the lock; that
run reports the outcome. After fixing a failure, clear the unit with
systemctl --user reset-failed chesstrainer-backup-pull.service (or
sudo systemctl reset-failed chesstrainer-backup.service).
Enabling the pull timer fires it once straight away: its OnBootSec=10m is
measured from boot, so enabling it hours after boot triggers immediately. That
is expected and harmless.
Monitoring
systemctl list-timers chesstrainer-backup.timer
sudo journalctl -u chesstrainer-backup --since "2 days ago" -o cat
sudo systemctl --failed
systemctl --user list-timers chesstrainer-backup-pull.timer chesstrainer-backup-drill.timer
journalctl --user -u chesstrainer-backup-pull --since "2 days ago" -o cat
A failed run produces both backup.failed (from the script) and
backup.unit-failed (from the OnFailure unit). The healthchecks.io check is
the backstop: if the VPS dies, the ping stops and you are emailed. The pull
fails when the newest archive is more than 48 hours old, which catches a stopped
VPS timer the next time the workstation pulls.
Troubleshooting
| Symptom | Likely cause |
|---|---|
backup.failed: missing command: age |
sudo apt install age |
backup.failed: only N MB free |
disk full on the VPS; prune old releases or grow the disk |
backup.failed: pg_dump failed |
chesstrainer-db is down or restarting — check docker ps |
backup.failed: globals dump is missing role … |
the role was renamed; update EXPECTED_ROLES |
pull.failed: rsync … failed |
SSH key, host key, or the VPS is unreachable |
pull.failed: newest archive … is Nh old |
the VPS backup timer stopped; check it before touching the workstation |
drill.failed: pg_restore failed |
archive is corrupt; keep it for diagnosis and restore an earlier one |
Never put dumps anywhere under priv/static or the release directory, and never
restore over chesstrainer_prod while the application is running.
New VPS checklist
- Ubuntu 24.04, DNS pointed at the VPS, firewall allows 22/80/443.
- Docker, Compose, Caddy, and runtime libraries installed.
chesstraineruser,/opt/chesstrainerlayout, root-owned env files.chesstrainer-dbmerged into the unified compose and healthy.- systemd unit installed and enabled;
ConditionPathExistskeeps it skipped until a release exists. - Release built on a matching Ubuntu runner and activated.
/healthand/verified on127.0.0.1:4100.- Caddy block,
cloudflare-ranges.caddy, and the CIDR refresh timer added; validated and reloaded. - Public
/,/privacy,/terms-conditions, and/healthverified. - Read-only role created and tunnel verified.
ageinstalled,/etc/chesstrainer/backup.envwritten, backup script and units installed, timer enabled, and one manual run produces a dump plus a globals file.- Workstation age key and pull configuration installed, user timers and linger enabled, first pull verified, first drill passes.
/etc/chesstrainer/probe.envwritten, probe script and units installed, timer enabled, and the healthchecks.io check green within one period.- Remote shell attaches and exits cleanly:
sudo /opt/chesstrainer/current/bin/chesstrainer remote(Ctrl-C twice).