fix(companion): Only reload companion manager when its config changed

A SIGHUP web reload recycles HTTP workers and re-reads config, but with
--preload it does not re-import application code: the WSGI callable is
loaded once and cached, so running companions are already current.
Restarting the manager on every reload bounced all companions for
nothing, slowing the common fast-reload path Frappe relies on.

reload_companion_manager now rebuilds the companion configs and compares
their sorted config_hash against the running manager's. Unchanged ->
leave the manager and its companions running. Changed (field, added, or
removed name) -> restart the manager from the fresh cfg, as before.
Per-companion reread via the control socket is unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/design/companion-process-manager.md

@@ -689,7 +689,7 @@ No per-companion logic in Arbiter.
 - [x] Implement transactional `reread`.
 - [x] Add manager spawn/reap logic in Arbiter.
 - [x] Add manager shutdown handling in Arbiter.
-- [x] Wire Gunicorn reload to manager `reread` or restart.
+- [x] Wire Gunicorn reload to restart the manager only when companion config changed.
 - [x] Close Gunicorn-only fds in manager child.
 - [x] Close manager-only fds in companion child.
 - [x] Add parent-death cleanup.

gunicorn/arbiter.py

@@ -735,20 +735,47 @@ class Arbiter:
             worker.tmp.close()
 
     def reload_companion_manager(self):
-        """Restart the companion manager so it picks up the new configuration.
+        """Reconcile the companion manager with the reloaded configuration.
 
-        Gunicorn reload (SIGHUP) rebuilds cfg. A running manager is asked to
-        stop -- it drains its companions first -- and the SIGCHLD reaper then
-        clears its pid so manage_companion_manager respawns it from the fresh
-        cfg. If companions were added where none ran, a new manager starts
-        right away. Per-companion transactional reread stays available
-        separately through the control socket.
+        A web reload (SIGHUP) recycles HTTP workers and re-reads config, but
+        with ``--preload`` it does not reload application code -- the WSGI
+        callable is loaded once and cached -- so the running companions are
+        already current. Restarting them on every reload would bounce them for
+        nothing, so the manager is only reloaded when the companion config
+        actually changed.
+
+        When it did change, a running manager is asked to stop (it drains its
+        companions first); the SIGCHLD reaper clears its pid so the main loop
+        respawns it from the fresh cfg, and a manager started where none ran
+        comes up right away. An unchanged config leaves the manager and its
+        companions untouched. Per-companion transactional reread stays
+        available separately through the control socket.
         """
+        try:
+            new_configs = build_companion_configs(self.cfg)
+        except Exception:
+            self.log.exception(
+                "Could not read companion config on reload; "
+                "leaving companion manager unchanged")
+            return
+        if not self._companion_configs_changed(new_configs):
+            return
         if self.companion_manager_pid != 0:
-            self.log.info("Reloading companion manager")
+            self.log.info("Companion config changed, reloading companion manager")
             self.stop_companion_manager(signal.SIGTERM)
         self.manage_companion_manager()
 
+    def _companion_configs_changed(self, new_configs):
+        """True when the companion config differs from the running manager's.
+
+        Compares the sorted ``config_hash`` of every companion, so a changed
+        field, an added name, or a removed name all count, while a pure web
+        reload with the same companion specs does not.
+        """
+        old_hashes = sorted(c.config_hash for c in self._companion_configs)
+        new_hashes = sorted(c.config_hash for c in new_configs)
+        return old_hashes != new_hashes
+
     def stop_companion_manager(self, sig):
         """Signal the companion manager to exit, if it is running.
 

gunicorn/companion/README.md

@@ -121,8 +121,17 @@ echo '{"cmd": "status"}' | socat - UNIX-CONNECT:/run/gunicorn/companion.sock
 Commands: `status`, `start <name>`, `stop <name>`, `restart <name>`, `reread`.
 
 `reread` is transactional: the new config is validated first, and on any error
-nothing changes and the old config keeps running. A `SIGHUP` to Gunicorn
-restarts the manager with the reloaded config.
+nothing changes and the old config keeps running.
+
+A `SIGHUP` to Gunicorn recycles the HTTP workers, then reloads the companion
+manager **only if the companion config changed**. With `--preload`, a reload
+does not re-import application code (the WSGI callable is loaded once and
+cached), so unchanged companions are already current and are left running --
+the common web reload never bounces them. When the companion specs do change,
+the manager is restarted with the new config. Note that, like the HTTP
+workers, companions pick up new application code only on a full restart, not on
+`SIGHUP`. Use `reread` for finer, per-companion changes without touching the
+others.
 
 ## Files
 

tests/test_arbiter.py

@@ -9,6 +9,7 @@ from unittest import mock
 
 import gunicorn.app.base
 import gunicorn.arbiter
+from gunicorn.companion.config import build_companion_configs
 from gunicorn.config import ReusePort
 
 
@@ -272,6 +273,35 @@ def test_reload_companion_manager_starts_when_none_running():
     arbiter.spawn_companion_manager.assert_called_once_with()
 
 
+def test_reload_companion_manager_noop_when_config_unchanged():
+    arbiter = gunicorn.arbiter.Arbiter(DummyApplication())
+    arbiter.cfg.set("companion_workers", [{"name": "rq", "target": "pkg:run"}])
+    # The running manager was spawned with this exact config.
+    arbiter._companion_configs = build_companion_configs(arbiter.cfg)
+    arbiter.companion_manager_pid = 4242
+    arbiter.stop_companion_manager = mock.Mock()
+    arbiter.spawn_companion_manager = mock.Mock()
+    arbiter.reload_companion_manager()
+    # A web reload with unchanged companion specs leaves the manager alone.
+    arbiter.stop_companion_manager.assert_not_called()
+    arbiter.spawn_companion_manager.assert_not_called()
+
+
+def test_reload_companion_manager_restarts_when_field_changed():
+    arbiter = gunicorn.arbiter.Arbiter(DummyApplication())
+    arbiter.cfg.set("companion_workers",
+                    [{"name": "rq", "target": "pkg:run", "startsecs": 1}])
+    arbiter._companion_configs = build_companion_configs(arbiter.cfg)
+    arbiter.companion_manager_pid = 4242
+    # Same name, changed field -> different config_hash -> reload.
+    arbiter.cfg.set("companion_workers",
+                    [{"name": "rq", "target": "pkg:run", "startsecs": 9}])
+    arbiter.stop_companion_manager = mock.Mock()
+    arbiter.spawn_companion_manager = mock.Mock()
+    arbiter.reload_companion_manager()
+    arbiter.stop_companion_manager.assert_called_once_with(signal.SIGTERM)
+
+
 @mock.patch('gunicorn.sock.close_sockets')
 def test_arbiter_stop_signals_companion_manager(close_sockets):
     arbiter = gunicorn.arbiter.Arbiter(DummyApplication())