A couple of replication fixes for device lists

changelog.d/4109.misc

@@ -0,0 +1 @@
+Reduce replication traffic for device lists

synapse/replication/tcp/protocol.py

@@ -656,7 +656,7 @@ def transport_kernel_read_buffer_size(protocol, read=True):
     "",
     ["command", "name"],
     lambda: {
-        (k[0], p.name,): count
+        (k, p.name,): count
         for p in connected_connections
         for k, count in iteritems(p.inbound_commands_counter)
     },
@@ -667,7 +667,7 @@ def transport_kernel_read_buffer_size(protocol, read=True):
     "",
     ["command", "name"],
     lambda: {
-        (k[0], p.name,): count
+        (k, p.name,): count
         for p in connected_connections
         for k, count in iteritems(p.outbound_commands_counter)
     },

synapse/storage/devices.py

@@ -589,10 +589,14 @@ def get_all_device_list_changes_for_remotes(self, from_key, to_key):
         combined list of changes to devices, and which destinations need to be
         poked. `destination` may be None if no destinations need to be poked.
         """
+        # We do a group by here as there can be a large number of duplicate
+        # entries, since we throw away device IDs.
         sql = """
-            SELECT stream_id, user_id, destination FROM device_lists_stream
+            SELECT MAX(stream_id) AS stream_id, user_id, destination
+            FROM device_lists_stream
             LEFT JOIN device_lists_outbound_pokes USING (stream_id, user_id, device_id)
             WHERE ? < stream_id AND stream_id <= ?
+            GROUP BY user_id, destination
         """
         return self._execute(
             "get_all_device_list_changes_for_remotes", None,