Improve performance of delete device messages query

changelog.d/16483.misc

@@ -0,0 +1 @@
+Improve performance of delete device messages query, cf issue [16479](https://github.com/matrix-org/synapse/issues/16479).

synapse/handlers/device.py

@@ -601,6 +601,8 @@ async def delete_devices(self, user_id: str, device_ids: List[str]) -> None:
                 )
 
             # Delete device messages asynchronously and in batches using the task scheduler
+            # We specify an upper stream id to avoid deleting non delivered messages
+            # if an user re-uses a device ID.
             await self._task_scheduler.schedule_task(
                 DELETE_DEVICE_MSGS_TASK_NAME,
                 resource_id=device_id,

synapse/storage/databases/main/deviceinbox.py

@@ -478,17 +478,18 @@ async def delete_messages_for_device(
                 log_kv({"message": "No changes in cache since last check"})
                 return 0
 
-        ROW_ID_NAME = self.database_engine.row_id_name
-
         def delete_messages_for_device_txn(txn: LoggingTransaction) -> int:
             sql = f"""
-                DELETE FROM device_inbox WHERE {ROW_ID_NAME} IN (
-                  SELECT {ROW_ID_NAME} FROM device_inbox
-                  WHERE user_id = ? AND device_id = ? AND stream_id <= ?
-                  LIMIT {limit}
+                DELETE FROM device_inbox WHERE user_id = ? AND device_id = ? AND stream_id <= (
+                  SELECT MAX(stream_id) FROM (
+                    SELECT stream_id FROM device_inbox
+                    WHERE user_id = ? AND device_id = ? AND stream_id <= ?
+                    ORDER BY stream_id
+                    LIMIT {limit}
+                  ) AS q1
                 )
                 """
-            txn.execute(sql, (user_id, device_id, up_to_stream_id))
+            txn.execute(sql, (user_id, device_id, user_id, device_id, up_to_stream_id))
             return txn.rowcount
 
         count = await self.db_pool.runInteraction(