Fix losing initial payload because of the race.

Author: Andrey Zelenchuk

channels_graphql_ws/graphql_ws_consumer.py

@@ -713,9 +713,6 @@ async def _register_subscription(
         # `_sids_by_group` without any locks.
         self._assert_thread()
 
-        # The subject we will trigger on the `broadcast` message.
-        trigger = rx.subjects.Subject()
-
         # The subscription notification queue.
         queue_size = notification_queue_limit
         if not queue_size or queue_size < 0:
@@ -728,56 +725,41 @@ async def _register_subscription(
 
         # Start an endless task which listens the `notification_queue`
         # and invokes subscription "resolver" on new notifications.
-        async def notifier():
+        async def notifier(observer: rx.Observer):
             """Watch the notification queue and notify clients."""
 
             # Assert we run in a proper thread.
             self._assert_thread()
-
-            # Dirty hack to partially workaround the race between:
-            #  1) call to `result.subscribe` in `_on_gql_start`; and
-            #  2) call to `trigger.on_next` below in this function.
-            # The first call must be earlier. Otherwise, first one or more notifications
-            # may be lost.
-            await asyncio.sleep(1)
-
             while True:
                 serialized_payload = await notification_queue.get()
 
                 # Run a subscription's `publish` method (invoked by the
-                # `trigger.on_next` function) within the threadpool used
+                # `observer.on_next` function) within the threadpool used
                 # for processing other GraphQL resolver functions.
                 # NOTE: it is important to run the deserialization
                 # in the worker thread as well.
                 def workload():
                     try:
                         payload = Serializer.deserialize(serialized_payload)
                     except Exception as ex:  # pylint: disable=broad-except
-                        trigger.on_error(f"Cannot deserialize payload. {ex}")
+                        observer.on_error(f"Cannot deserialize payload. {ex}")
                     else:
-                        trigger.on_next(payload)
+                        observer.on_next(payload)
 
                 await self._run_in_worker(workload)
 
                 # Message processed. This allows `Queue.join` to work.
                 notification_queue.task_done()
 
-        # Enqueue the `publish` method execution. But do not notify
-        # clients when `publish` returns `SKIP`.
-        stream = trigger.map(publish_callback).filter(  # pylint: disable=no-member
-            lambda publish_returned: publish_returned is not self.SKIP
-        )
-
+        def push_payloads(observer: rx.Observer):
             # Start listening for broadcasts (subscribe to the Channels
             # groups), spawn the notification processing task and put
             # subscription information into the registry.
             # NOTE: Update of `_sids_by_group` & `_subscriptions` must be
             # atomic i.e. without `awaits` in between.
-        waitlist = []
             for group in groups:
                 self._sids_by_group.setdefault(group, []).append(operation_id)
-            waitlist.append(self._channel_layer.group_add(group, self.channel_name))
-        notifier_task = self._spawn_background_task(notifier())
+            notifier_task = self._spawn_background_task(notifier(observer))
             self._subscriptions[operation_id] = self._SubInf(
                 groups=groups,
                 sid=operation_id,
@@ -786,9 +768,20 @@ def workload():
                 notifier_task=notifier_task,
             )
 
-        await asyncio.wait(waitlist)
+        await asyncio.wait(
+            [
+                self._channel_layer.group_add(group, self.channel_name)
+                for group in groups
+            ]
+        )
 
-        return stream
+        # Enqueue the `publish` method execution. But do not notify
+        # clients when `publish` returns `SKIP`.
+        return (
+            rx.Observable.create(push_payloads)  # pylint: disable=no-member
+            .map(publish_callback)
+            .filter(lambda publish_returned: publish_returned is not self.SKIP)
+        )
 
     async def _on_gql_stop(self, operation_id):
         """Process the STOP message.