Use READ COMMITTED isolation level when purging rooms

changelog.d/12942.misc

@@ -0,0 +1 @@
+Use lower isolation level when purging rooms to avoid serialization errors. Contributed by Nick @ Beeper.

synapse/storage/databases/main/purge_events.py

@@ -19,6 +19,7 @@
 from synapse.storage.database import LoggingTransaction
 from synapse.storage.databases.main import CacheInvalidationWorkerStore
 from synapse.storage.databases.main.state import StateGroupWorkerStore
+from synapse.storage.engines._base import IsolationLevel
 from synapse.types import RoomStreamToken
 
 logger = logging.getLogger(__name__)
@@ -318,10 +319,21 @@ async def purge_room(self, room_id: str) -> List[int]:
             The list of state groups to delete.
         """
         return await self.db_pool.runInteraction(
-            "purge_room", self._purge_room_txn, room_id
+            "purge_room",
+            self._purge_room_txn,
+            room_id=room_id,
+            # This is safe because we don't care if room data is updated during the transaction
+            isolation_level=IsolationLevel.READ_COMMITTED,
         )
 
     def _purge_room_txn(self, txn: LoggingTransaction, room_id: str) -> List[int]:
+        # This collides with event persistence so we cannot write new events and metadata into
+        # a room while deleting it or this transaction will fail.
+        txn.execute(
+            "SELECT room_version FROM rooms WHERE room_id = ? FOR UPDATE",
+            (room_id,),
+        )
+
         # First, fetch all the state groups that should be deleted, before
         # we delete that information.
         txn.execute(