Prioritize outbound to-device over device list updates (#13922)

Otherwise device list changes for large accounts can temporarily delay to-device messages.

2 files changed, 17 insertions, 13 deletions

diff --git a/changelog.d/13922.bugfix b/changelog.d/13922.bugfix
new file mode 100644
index 0000000000..7269d28dee
--- /dev/null
+++ b/changelog.d/13922.bugfix
@@ -0,0 +1 @@
+Fix long-standing bug where device updates could cause delays sending out to-device messages over federation.
diff --git a/synapse/federation/sender/per_destination_queue.py b/synapse/federation/sender/per_destination_queue.py
index 41d8b937af..084c45a95c 100644
--- a/synapse/federation/sender/per_destination_queue.py
+++ b/synapse/federation/sender/per_destination_queue.py
@@ -646,29 +646,32 @@ class _TransactionQueueManager:
 
         # We start by fetching device related EDUs, i.e device updates and to
         # device messages. We have to keep 2 free slots for presence and rr_edus.
-        limit = MAX_EDUS_PER_TRANSACTION - 2
-
-        device_update_edus, dev_list_id = await self.queue._get_device_update_edus(
-            limit
-        )
-
-        if device_update_edus:
-            self._device_list_id = dev_list_id
-        else:
-            self.queue._last_device_list_stream_id = dev_list_id
-
-        limit -= len(device_update_edus)
+        device_edu_limit = MAX_EDUS_PER_TRANSACTION - 2
 
+        # We prioritize to-device messages so that existing encryption channels
+        # work. We also keep a few slots spare (by reducing the limit) so that
+        # we can still trickle out some device list updates.
         (
             to_device_edus,
             device_stream_id,
-        ) = await self.queue._get_to_device_message_edus(limit)
+        ) = await self.queue._get_to_device_message_edus(device_edu_limit - 10)
 
         if to_device_edus:
             self._device_stream_id = device_stream_id
         else:
             self.queue._last_device_stream_id = device_stream_id
 
+        device_edu_limit -= len(to_device_edus)
+
+        device_update_edus, dev_list_id = await self.queue._get_device_update_edus(
+            device_edu_limit
+        )
+
+        if device_update_edus:
+            self._device_list_id = dev_list_id
+        else:
+            self.queue._last_device_list_stream_id = dev_list_id
+
         pending_edus = device_update_edus + to_device_edus
 
         # Now add the read receipt EDU.