@Falco @satonotdead
Logré solucionarlo. Esperé bastante tiempo para asegurarme de que estuviera realmente arreglado, y parece que sí.
La configuración
Discourse se ejecuta en Docker y expone nginx a través de un socket Unix (/var/discourse/shared/standalone/nginx.http.sock). Caddy se encuentra delante de él como proxy inverso y transmite la IP real del cliente a través de X-Real-IP.
Hubo que cambiar dos cosas: la configuración de nginx dentro del contenedor y la forma en que ejecuto la reconstrucción.
1. app.yml
## Plugins go here
## see https://meta.discourse.org/t/19157 for details
hooks:
# This after_code block is just my own plugin list — it has nothing to do with
# the fix. Keep whatever you already have here.
after_code:
- exec:
[...]
# This is the part that matters.
# 1. Writes an http-level map that turns the literal "unix:" into 127.0.0.1.
# The 00- prefix makes nginx load it before discourse.conf.
# 2. Rewrites discourse.conf so X-Forwarded-For uses the mapped variable
# instead of the raw $remote_addr.
# 3. nginx -t fails the build if the result is not valid.
after_web_config:
- exec: >-
printf 'map $remote_addr $safe_remote_addr {\n "unix:" 127.0.0.1;\n default $remote_addr;\n}\nreal_ip_header X-Real-IP;\n'
> /etc/nginx/conf.d/00-safe-remote-addr.conf
- exec: >-
sed -i 's/X-Forwarded-For \$remote_addr;/X-Forwarded-For $safe_remote_addr;/g'
/etc/nginx/conf.d/discourse.conf
- exec: nginx -t
El código crítico es lo que se encuentra dentro del cuerpo de after_web_config.
2. El bloque Caddyfile
El siguiente script gestiona el modo de mantenimiento a través de un archivo de bandera, por lo que tu bloque de sitio debe conocerlo:
forum.example.com {
request_header X-Real-IP {remote_host}
root * /var/caddy/flags
@maintenance file maintenance.flag
handle @maintenance {
respond "Maintenance in progress. We'll be back in a few minutes." 503
}
handle {
reverse_proxy unix//var/discourse/shared/standalone/nginx.http.sock
}
}
La ruta de la bandera en el script y el archivo que busca el matcher @maintenance deben ser el mismo archivo. Si renombras uno, renombra el otro, de lo contrario el modo de mantenimiento nunca se activará silenciosamente.
Si la misma instancia de Caddy también sirve otras aplicaciones, usa una bandera que solo coincida con el bloque del foro, de lo contrario una reconstrucción de Discourse tumbará también esas otras aplicaciones.
3. El script de reconstrucción
Ya no uso ./launcher rebuild app directamente. Uso discourse-rebuild.sh:
#!/bin/bash
#
# discourse-rebuild.sh — Rebuilds the Discourse container without leaving orphaned
# requests behind and without the `invalid input syntax for type inet: "unix:"` error.
#
# CONTEXT
# Discourse runs in Docker and exposes nginx on a Unix socket
# (/var/discourse/shared/standalone/nginx.http.sock). Caddy acts as a reverse
# proxy in front of it and passes the client's real IP through X-Real-IP.
#
# A `rebuild` destroys the container and recreates the socket with a new inode.
# During that transition there are two problematic windows:
#
# 1) Caddy keeps state from the previous socket until it is reloaded.
# 2) nginx starts accepting connections as soon as it boots, but Unicorn
# takes ~15s longer before it can serve them.
#
# A request landing in either window may arrive with no X-Real-IP. $remote_addr
# is then left holding the literal "unix:", which PostgreSQL rejects when
# inserting it into an inet column -> HTTP 500.
#
# WHAT IT DOES
# 1. Raises a flag file that puts the site into 503 (Caddy checks it on every
# request, so it takes effect instantly and with no reload).
# 2. Waits until nothing but nginx itself is holding the socket open.
# 3. Rebuilds the container.
# 4. Polls /srv/status against the socket until Unicorn answers 200.
# 5. Reloads Caddy so it picks up the new socket, and clears the flag.
# 6. Starts the watcher that logs any leftover "unix:" hit.
#
# If the rebuild fails, or Discourse never answers within the polling window,
# the flag is NOT removed: the site stays in maintenance on purpose, so a broken
# container is never exposed. Bring it back up by hand with:
# rm -f /var/caddy/flags/maintenance.flag
#
# REQUIREMENTS
# - lsof installed, and root privileges.
# - FLAG below must point at the exact same file the Caddyfile @maintenance
# matcher looks for.
# - request_header X-Real-IP {remote_host} in that same Caddyfile block.
#
# USAGE
# ./discourse-rebuild.sh
#
# TO SEE WHAT THE WATCHER CAUGHT
# cat /var/log/unixip-hits.log
#
set -e
FLAG=/var/caddy/flags/maintenance.flag
SOCK=/var/discourse/shared/standalone/nginx.http.sock
cleanup() {
local code=$?
systemctl reload caddy
if [ $code -eq 0 ]; then
rm -f "$FLAG"
echo "✅ Rebuild finished. Site is online."
else
echo "⚠️ Rebuild failed (exit code $code). The site is still in maintenance."
echo " Check it, and once it's ready: rm -f $FLAG"
fi
}
trap cleanup EXIT
mkdir -p "$(dirname "$FLAG")"
touch "$FLAG"
echo "🔧 Maintenance is on. Waiting for in-flight requests..."
# Rough heuristic: count the processes holding the socket open and wait until
# only the listener is left. Up to 30s.
for i in $(seq 30); do
n=$(lsof -t "$SOCK" 2>/dev/null | wc -l || echo 0)
[ "$n" -le 1 ] && break
sleep 1
done
/var/discourse/launcher rebuild app
# Up to 60 attempts: about 2 minutes of sleeps, more if any curl hits its own
# 5s timeout.
echo "⏳ Waiting for Discourse to answer..."
status=000
for i in $(seq 60); do
status=$(curl -s -o /dev/null -w '%{http_code}' --max-time 5 \
--unix-socket "$SOCK" http://localhost/srv/status 2>/dev/null || echo 000)
[ "$status" = "200" ] && break
sleep 2
done
if [ "$status" != "200" ]; then
echo "⚠️ Discourse never answered within the polling window (last status code: $status)."
exit 1
fi
echo "✅ Discourse ready after ~$((i*2))s."
systemd-run --unit=unixip-watch --collect \
/bin/bash -c "docker exec app tail -F /var/log/nginx/access.log | grep --line-buffered 'unix:' >> /var/log/unixip-hits.log"
Lo que hace realmente el script
- Pone el sitio en mantenimiento a través del archivo de bandera de Caddy.
- Espera a que el socket se calme y luego ejecuta
./launcher rebuild app.
- Realiza consultas a
/srv/status a través del socket hasta que Unicorn responda 200. Esta es la parte crítica: es lo que impide que las solicitudes lleguen a nginx mientras Unicorn aún se está iniciando.
- Recarga Caddy (por si acaso, resultó no ser crítico).
- Borra la bandera de mantenimiento solo si esa consulta tuvo éxito.
Después de varias semanas y varias reconstrucciones, no he vuelto a ver ese error.
PD: De verdad no sé qué demonios hice.