GELF `source` is the literal `{{.Container.Config.Hostname}}` since add-on 1.12.0 (hostname resolved at package init, before the launcher sets SYSLOG_HOSTNAME)

[john120283]

Summary

Since the Logspout Home Assistant add-on 1.12.0 (the version that "replaced the old shell-based startup wrapper with a native Home Assistant launcher"), every log forwarded to Graylog arrives with its source set to the literal, unrendered string:

{{.Container.Config.Hostname}}

instead of the configured hostname (e.g. homeassistant). The previous add-on version 1.11.0 works correctly.

• Symptom is observed in: bertbaron/hassio-addons, add-on logspout, version 1.12.0.
• Root cause / code fix belongs in: bertbaron/logspout (the add-on Dockerfile builds this repo via LOGSPOUT_SOURCE=https://github.com/bertbaron/logspout.git, building ./cmd/launcher).

User-visible symptom

In Graylog, the source field (which comes from the GELF host field) is the literal template string {{.Container.Config.Hostname}} for all messages, rather than the host name set via the add-on's hostname: homeassistant option. This makes it impossible to filter/group logs by source.

Root cause: init-order vs os.Setenv

In adapters/gelf/gelf.go, the GELF host was resolved at package init() time:

var hostname string

func getHostname() string {
    content, err := os.ReadFile("/etc/host_hostname")
    if err == nil && len(content) > 0 {
        hostname = strings.TrimRight(string(content), "\r\n")
    } else {
        hostname = cfg.GetEnvDefault("SYSLOG_HOSTNAME", "{{.Container.Config.Hostname}}")
    }
    return hostname
}

func init() {
    hostname = getHostname()                       // runs at package init, before main()
    router.AdapterFactories.Register(NewGelfAdapter, "gelf")
}

The native launcher (launcher/launcher.go, RunWithRunner) reads the add-on hostname option and applies it to the environment inside main():

env := BuildEnvironment(config, ...) // includes "SYSLOG_HOSTNAME": config.Hostname
...
for key, value := range env {
    os.Setenv(key, value)   // happens in main(), AFTER gelf's init() already read SYSLOG_HOSTNAME
}

Go runs all package init() functions before main(). So the ordering is:

1. The gelf package init() runs getHostname().
   • SYSLOG_HOSTNAME is not yet set (the launcher sets it later in main()).
   • /etc/host_hostname no longer exists — the old shell wrapper that used to create/export this was removed in the 1.12.0 rewrite.
   • So it falls back to the literal default {{.Container.Config.Hostname}}.
2. main() runs the launcher, which calls os.Setenv("SYSLOG_HOSTNAME", "homeassistant") — but it is too late, the package-level hostname variable was already frozen to the literal default.

In 1.11.0 the bash wrapper exported SYSLOG_HOSTNAME (and/or wrote /etc/host_hostname) before the binary started, so the gelf init() saw the correct value. The Go-launcher rewrite moved that work from "before process start" to "inside main()", which is after package init — introducing this regression.

Why config / env workarounds don't help

You cannot work around this from the add-on configuration:

• Setting hostname: homeassistant is exactly what's broken — the launcher does set SYSLOG_HOSTNAME from it, but only in main(), after init() already ran.
• Adding SYSLOG_HOSTNAME via the add-on's custom env list does not help either: the launcher applies those with the same os.Setenv loop in main(), again after init(). (And SYSLOG_HOSTNAME is a reserved managed key in the launcher anyway.)

The only correct fix is to stop resolving the hostname at package init() time.

Suggested fix

Resolve the hostname lazily — after the launcher's os.Setenv has run — rather than at package init(). For example, resolve it inside NewGelfAdapter (called when routes are set up, after main()), keeping the existing precedence /etc/host_hostname -> SYSLOG_HOSTNAME -> literal default, and keeping the init() adapter-factory registration.

Environment

• Logspout HA add-on (bertbaron/hassio-addons, logspout/): 1.12.0 (broken), 1.11.0 (works).
• Forwarder: Graylog GELF (UDP/TCP/TLS).

Proposed fix (implemented & tested)

I have a minimal, behavior-preserving patch that makes getHostname() a pure function and resolves the hostname in NewGelfAdapter (per-adapter, after the launcher's os.Setenv) instead of at package init(). Lookup precedence (/etc/host_hostname → SYSLOG_HOSTNAME → literal default) is unchanged.

--- a/adapters/gelf/gelf.go
+++ b/adapters/gelf/gelf.go
@@
-var hostname string
-
+// getHostname resolves the GELF host field, preferring the /etc/host_hostname
+// file, then the SYSLOG_HOSTNAME environment variable, then a literal default.
+// It must be called lazily (e.g. when an adapter is created) rather than at
+// package init() time, so that any launcher that sets SYSLOG_HOSTNAME via
+// os.Setenv during main() has already run.
 func getHostname() string {
 	content, err := os.ReadFile("/etc/host_hostname")
 	if err == nil && len(content) > 0 {
-		hostname = strings.TrimRight(string(content), "\r\n")
-	} else {
-		hostname = cfg.GetEnvDefault("SYSLOG_HOSTNAME", "{{.Container.Config.Hostname}}")
+		return strings.TrimRight(string(content), "\r\n")
 	}
-	return hostname
+	return cfg.GetEnvDefault("SYSLOG_HOSTNAME", "{{.Container.Config.Hostname}}")
 }
 
 func init() {
-	hostname = getHostname()
 	router.AdapterFactories.Register(NewGelfAdapter, "gelf")
 }
 
 // Adapter is an adapter that streams UDP JSON to Graylog
 type Adapter struct {
-	writer gelf.Writer
-	route  *router.Route
+	writer   gelf.Writer
+	route    *router.Route
+	hostname string
 }
 
 // NewGelfAdapter creates an Adapter with UDP as the default transport.
@@ func NewGelfAdapter(route *router.Route) (router.LogAdapter, error) {
 	return &Adapter{
-		route:  route,
-		writer: gelfWriter,
+		route:    route,
+		writer:   gelfWriter,
+		hostname: getHostname(),
 	}, nil
 }
@@ func (a *Adapter) Stream(logstream chan *router.Message) {
 		msg := gelf.Message{
 			Version:  "1.1",
-			Host:     hostname,
+			Host:     a.hostname,

Plus two regression tests in adapters/gelf/gelf_test.go that set SYSLOG_HOSTNAME after import and assert it is picked up (these fail against the old init()-based code). go test ./... passes; gofmt and go vet are clean.

Happy to open a PR for this if you're open to it.

[bertbaron]

Thanks for posting the issue. The same issue also appeared in the Loki adapter, so I applied the same fix to both, see this commit

I will create a new release this weekend.