Fix gunicorn endless restarting (#12879)

CHANGES/12879.bugfix.rst

@@ -0,0 +1 @@
+Fixed ``GunicornWebWorker`` endlessly reloading when app fails during startup -- by :user:`Dreamsorcerer`.

aiohttp/worker.py

@@ -60,14 +60,20 @@ def init_process(self) -> None:
         super().init_process()
 
     def run(self) -> None:
-        self._task = self.loop.create_task(self._run())
+        # base.Worker.init_process() sets self.booted = True before
+        # invoking run(), but for the aiohttp worker the real boot work
+        # (factory call, runner setup, binding sockets) happens here.
+        # Reset until _run() reaches the serve loop so that the arbiter
+        # can tell a startup failure from a normal worker exit and
+        # halt instead of endlessly respawning workers.
+        self.booted = False
 
-        try:  # ignore all finalization problems
+        self._task = self.loop.create_task(self._run())
+        try:
             self.loop.run_until_complete(self._task)
-        except Exception:
-            self.log.exception("Exception in gunicorn worker")
-        self.loop.run_until_complete(self.loop.shutdown_asyncgens())
-        self.loop.close()
+        finally:
+            self.loop.run_until_complete(self.loop.shutdown_asyncgens())
+            self.loop.close()
 
         sys.exit(self.exit_code)
 
@@ -118,6 +124,12 @@ async def _run(self) -> None:
             )
             await site.start()
 
+        # Sockets are bound; tell the arbiter the worker is ready to
+        # accept requests. Any failure before this point propagates out
+        # of run() with self.booted=False so the arbiter exits with
+        # WORKER_BOOT_ERROR instead of treating this as a clean exit.
+        self.booted = True
+
         # If our parent changed then we shut down.
         pid = os.getpid()
         try:

tests/test_worker.py

@@ -143,9 +143,33 @@ def test_run_not_app(
     worker.loop = loop
     worker.wsgi = "not-app"
     worker.alive = False
-    with pytest.raises(SystemExit):
+    with pytest.raises(RuntimeError, match="wsgi app should be"):
+        worker.run()
+    assert not worker.booted
+    assert loop.is_closed()
+
+
+def test_run_on_startup_raises(
+    worker: base_worker.GunicornWebWorker, loop: asyncio.AbstractEventLoop
+) -> None:
+    worker.log = mock.Mock()
+    worker.cfg = mock.Mock()
+    worker.cfg.access_log_format = ACCEPTABLE_LOG_FORMAT
+    worker.cfg.is_ssl = False
+    worker.cfg.graceful_timeout = 100
+    worker.sockets = []
+
+    app = web.Application()
+
+    async def boom(app: web.Application) -> None:
+        raise RuntimeError("boom during startup")
+
+    app.on_startup.append(boom)
+    worker.wsgi = app
+    worker.loop = loop
+    with pytest.raises(RuntimeError, match="boom during startup"):
         worker.run()
-    worker.log.exception.assert_called_with("Exception in gunicorn worker")
+    assert not worker.booted
     assert loop.is_closed()