Failing hosts no longer stop the run, but the script now counts them and exits 1 so callers notice the failure (the pr/4 rewrite already does this via regenerateExitCode).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every write of a host config now goes through one configure() that runs
under a re-entrant lock (lib/nginx-lock.js), so concurrent saves can't test
or rename each other's files:
- the new config is written and checked with nginx -t; when nginx rejects
this host, the previous file (and .err file) is put back and an
NginxConfigError with structured details is thrown (scope host/global,
file, line, humanised message and hint, raw nginx output)
- when another host's file breaks the test, that host is only blamed when it
fails on its own (lib/nginx-blame.js). It is regenerated from its database
row first and only moved to .err and marked offline when that fails too.
A conflict with the change being saved fails the save instead. An error
that names no file (eg. an unknown variable) is only a conflict when it
goes away with every other host's config moved aside; otherwise it is the
host's own error and, when its advanced config uses the variable or path,
reads "Advanced configuration: ..."
- errors are humanised (lib/nginx-errors.js) and mapped to the advanced
config of the host where possible (new begin/end markers in the templates)
- the API returns error.details; a middleware redacts details that name a
host the user can't see
- bulkGenerateConfigs (certificates) and the default site go through the
same path, a host nginx rejects is taken offline instead of failing all
- nginx -s reload only signals the master, which reads the config later.
reloadNginx() waits (bounded, 5 s) until the master runs a new generation
of workers and the old ones are shutting down (lib/nginx-reload.js, from
/proc) before the lock is released,
otherwise the next holder's rejected candidate file could be what the
master reads and the change saved before would never be applied. A reload
that can't be confirmed is repeated once at the end of the locked section,
after every rollback. When nginx logs an [emerg] for the reload (eg. a
listen port in use, which nginx -t doesn't catch) the wait ends as soon
as nginx gives up on it (for a bind() after its own retries), and after
a repeated reload that isn't confirmed either, later saves don't repeat
it until a reload is confirmed again, so they don't each wait 10 s
- reconcile() regenerates missing / moved-aside configs and retries hosts
whose upstream didn't resolve yet; the startup commit runs it from
backend/index.js (NPM_RECONCILE_ON_START)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
processItems() passed proxyHostModel and the literal "proxy_host" to
configure() for every host type. host_type selects both the template and
the output path, and each host type has its own id sequence, so
redirection hosts, 404 hosts and streams were rendered through
proxy_host.conf and written over /data/nginx/proxy_host/<id>.conf. The
proxy host sharing that id lost its config file and had the resulting
nginx error recorded in its own meta.
- Added a script to install every single plugin, used in development and debugging
- Improved certbot plugin install commands
- Adjusted some version for plugins to install properly
- It's noted that some plugins require deps that do not match other plugins,
however these use cases should be extremely rare