sketch of an idea to reduce cache misses for get_users_in_room for ASes

synapse/appservice/__init__.py

@@ -186,8 +186,13 @@ class ApplicationService(object):
 
     @cachedInlineCallbacks(num_args=1, cache_context=True)
     def _matches_user_in_member_list(self, room_id, store, cache_context):
+
+        def invalidate_only_when_as_users_change(room_id, member):
+            if self.is_interested_in_user(member):
+                cache_context.invalidate(room_id)
+
         member_list = yield store.get_users_in_room(
-            room_id, on_invalidate=cache_context.invalidate
+            room_id, on_invalidate=invalidate_only_when_as_users_change
         )
 
         # check joined member events

synapse/storage/events.py

@@ -793,6 +793,13 @@ class EventsStore(EventsWorkerStore):
                     self._invalidate_cache_and_stream(
                         txn, self.get_rooms_for_user_with_stream_ordering, (member,)
                     )
+                    # we call this once for each member to allow AS workers to
+                    # only invalidate their caches when AS members change.
+                    # other folks naively using a cached get_users_in_rooms will
+                    # just get prodded multiple times and can ignore the member key.
+                    self._invalidate_cache_and_stream(
+                        txn, self.get_users_in_room, (room_id, member)
+                    )
 
                 for host in set(get_domain_from_id(u) for u in members_changed):
                     self._invalidate_cache_and_stream(
@@ -802,10 +809,6 @@ class EventsStore(EventsWorkerStore):
                         txn, self.was_host_joined, (room_id, host)
                     )
 
-                self._invalidate_cache_and_stream(
-                    txn, self.get_users_in_room, (room_id,)
-                )
-
                 self._invalidate_cache_and_stream(
                     txn, self.get_current_state_ids, (room_id,)
                 )