Sliding Sync: Get bump_stamp from new sliding sync tables because i…

…t's faster (#17658)

Get `bump_stamp` from [new sliding sync
tables](#17512) which should
be faster (performance) than flipping through the latest events in the
room.

changelog.d/17658.misc

@@ -0,0 +1 @@
+Get `bump_stamp` from [new sliding sync tables](https://github.com/element-hq/synapse/pull/17512) which should be faster.

synapse/handlers/sliding_sync/__init__.py

@@ -1040,29 +1040,67 @@ async def get_room_sync_data(
                     )
                 )
 
-        # By default, just choose the membership event position
+        # Figure out the last bump event in the room
+        #
+        # By default, just choose the membership event position for any non-join membership
         bump_stamp = room_membership_for_user_at_to_token.event_pos.stream
-
-        # Figure out the last bump event in the room if we're in the room.
+        # If we're joined to the room, we need to find the last bump event before the
+        # `to_token`
         if room_membership_for_user_at_to_token.membership == Membership.JOIN:
-            last_bump_event_result = (
-                await self.store.get_last_event_pos_in_room_before_stream_ordering(
-                    room_id,
-                    to_token.room_key,
-                    event_types=SLIDING_SYNC_DEFAULT_BUMP_EVENT_TYPES,
-                )
+            # We can quickly query for the latest bump event in the room using the
+            # sliding sync tables.
+            latest_room_bump_stamp = await self.store.get_latest_bump_stamp_for_room(
+                room_id
             )
 
-            # But if we found a bump event, use that instead
-            if last_bump_event_result is not None:
-                _, new_bump_event_pos = last_bump_event_result
+            min_to_token_position = to_token.room_key.stream
 
-                # If we've just joined a remote room, then the last bump event may
-                # have been backfilled (and so have a negative stream ordering).
-                # These negative stream orderings can't sensibly be compared, so
-                # instead we use the membership event position.
-                if new_bump_event_pos.stream > 0:
-                    bump_stamp = new_bump_event_pos.stream
+            # If we can rely on the new sliding sync tables and the `bump_stamp` is
+            # `None`, just fallback to the membership event position. This can happen
+            # when we've just joined a remote room and all the events are backfilled.
+            if (
+                # FIXME: The background job check can be removed once we bump
+                # `SCHEMA_COMPAT_VERSION` and run the foreground update for
+                # `sliding_sync_joined_rooms`/`sliding_sync_membership_snapshots`
+                # (tracked by https://github.com/element-hq/synapse/issues/17623)
+                await self.store.have_finished_sliding_sync_background_jobs()
+                and latest_room_bump_stamp is None
+            ):
+                pass
+
+            # The `bump_stamp` stored in the database might be ahead of our token. Since
+            # `bump_stamp` is only a `stream_ordering` position, we can't be 100% sure
+            # that's before the `to_token` in all scenarios. The only scenario we can be
+            # sure of is if the `bump_stamp` is totally before the minimum position from
+            # the token.
+            #
+            # We don't need to check if the background update has finished, as if the
+            # returned bump stamp is not None then it must be up to date.
+            elif (
+                latest_room_bump_stamp is not None
+                and latest_room_bump_stamp < min_to_token_position
+            ):
+                bump_stamp = latest_room_bump_stamp
+
+            # Otherwise, if it's within or after the `to_token`, we need to find the
+            # last bump event before the `to_token`.
+            else:
+                last_bump_event_result = (
+                    await self.store.get_last_event_pos_in_room_before_stream_ordering(
+                        room_id,
+                        to_token.room_key,
+                        event_types=SLIDING_SYNC_DEFAULT_BUMP_EVENT_TYPES,
+                    )
+                )
+                if last_bump_event_result is not None:
+                    _, new_bump_event_pos = last_bump_event_result
+
+                    # If we've just joined a remote room, then the last bump event may
+                    # have been backfilled (and so have a negative stream ordering).
+                    # These negative stream orderings can't sensibly be compared, so
+                    # instead we use the membership event position.
+                    if new_bump_event_pos.stream > 0:
+                        bump_stamp = new_bump_event_pos.stream
 
         unstable_expanded_timeline = False
         prev_room_sync_config = previous_connection_state.room_configs.get(room_id)

synapse/storage/databases/main/events.py

@@ -327,6 +327,13 @@ async def _persist_events_and_state_updates(
 
         async with stream_ordering_manager as stream_orderings:
             for (event, _), stream in zip(events_and_contexts, stream_orderings):
+                # XXX: We can't rely on `stream_ordering`/`instance_name` being correct
+                # at this point. We could be working with events that were previously
+                # persisted as an `outlier` with one `stream_ordering` but are now being
+                # persisted again and de-outliered and are being assigned a different
+                # `stream_ordering` here that won't end up being used.
+                # `_update_outliers_txn()` will fix this discrepancy (always use the
+                # `stream_ordering` from the first time it was persisted).
                 event.internal_metadata.stream_ordering = stream
                 event.internal_metadata.instance_name = self._instance_name
 
@@ -470,11 +477,11 @@ async def _calculate_sliding_sync_table_changes(
                 membership_infos_to_insert_membership_snapshots.append(
                     # XXX: We don't use `SlidingSyncMembershipInfoWithEventPos` here
                     # because we're sourcing the event from `events_and_contexts`, we
-                    # can't rely on `stream_ordering`/`instance_name` being correct. We
-                    # could be working with events that were previously persisted as an
-                    # `outlier` with one `stream_ordering` but are now being persisted
-                    # again and de-outliered and assigned a different `stream_ordering`
-                    # that won't end up being used. Since we call
+                    # can't rely on `stream_ordering`/`instance_name` being correct at
+                    # this point. We could be working with events that were previously
+                    # persisted as an `outlier` with one `stream_ordering` but are now
+                    # being persisted again and de-outliered and assigned a different
+                    # `stream_ordering` that won't end up being used. Since we call
                     # `_calculate_sliding_sync_table_changes()` before
                     # `_update_outliers_txn()` which fixes this discrepancy (always use
                     # the `stream_ordering` from the first time it was persisted), we're
@@ -591,11 +598,17 @@ async def _calculate_sliding_sync_table_changes(
                         event_types=SLIDING_SYNC_DEFAULT_BUMP_EVENT_TYPES,
                     )
                 )
-                bump_stamp_to_fully_insert = (
-                    most_recent_bump_event_pos_results[1].stream
-                    if most_recent_bump_event_pos_results is not None
-                    else None
-                )
+                if most_recent_bump_event_pos_results is not None:
+                    _, new_bump_event_pos = most_recent_bump_event_pos_results
+
+                    # If we've just joined a remote room, then the last bump event may
+                    # have been backfilled (and so have a negative stream ordering).
+                    # These negative stream orderings can't sensibly be compared, so
+                    # instead just leave it as `None` in the table and we will use their
+                    # membership event position as the bump event position in the
+                    # Sliding Sync API.
+                    if new_bump_event_pos.stream > 0:
+                        bump_stamp_to_fully_insert = new_bump_event_pos.stream
 
                 current_state_ids_map = dict(
                     await self.store.get_partial_filtered_current_state_ids(
@@ -2123,31 +2136,26 @@ def _update_sliding_sync_tables_with_new_persisted_events_txn(
         if len(events_and_contexts) == 0:
             return
 
-        # We only update the sliding sync tables for non-backfilled events.
-        #
-        # Check if the first event is a backfilled event (with a negative
-        # `stream_ordering`). If one event is backfilled, we assume this whole batch was
-        # backfilled.
-        first_event_stream_ordering = events_and_contexts[0][
-            0
-        ].internal_metadata.stream_ordering
-        # This should exist for persisted events
-        assert first_event_stream_ordering is not None
-        if first_event_stream_ordering < 0:
-            return
-
         # Since the list is sorted ascending by `stream_ordering`, the last event should
         # have the highest `stream_ordering`.
         max_stream_ordering = events_and_contexts[-1][
             0
         ].internal_metadata.stream_ordering
+        # `stream_ordering` should be assigned for persisted events
+        assert max_stream_ordering is not None
+        # Check if the event is a backfilled event (with a negative `stream_ordering`).
+        # If one event is backfilled, we assume this whole batch was backfilled.
+        if max_stream_ordering < 0:
+            # We only update the sliding sync tables for non-backfilled events.
+            return
+
         max_bump_stamp = None
         for event, _ in reversed(events_and_contexts):
             # Sanity check that all events belong to the same room
             assert event.room_id == room_id
 
             if event.type in SLIDING_SYNC_DEFAULT_BUMP_EVENT_TYPES:
-                # This should exist for persisted events
+                # `stream_ordering` should be assigned for persisted events
                 assert event.internal_metadata.stream_ordering is not None
 
                 max_bump_stamp = event.internal_metadata.stream_ordering
@@ -2156,11 +2164,6 @@ def _update_sliding_sync_tables_with_new_persisted_events_txn(
                 # matching bump event which should have the highest `stream_ordering`.
                 break
 
-        # We should have exited earlier if there were no events
-        assert (
-            max_stream_ordering is not None
-        ), "Expected to have a stream_ordering if we have events"
-
         # Handle updating the `sliding_sync_joined_rooms` table.
         #
         txn.execute(

synapse/storage/databases/main/sliding_sync.py

@@ -41,6 +41,46 @@
 
 
 class SlidingSyncStore(SQLBaseStore):
+    async def get_latest_bump_stamp_for_room(
+        self,
+        room_id: str,
+    ) -> Optional[int]:
+        """
+        Get the `bump_stamp` for the room.
+
+        The `bump_stamp` is the `stream_ordering` of the last event according to the
+        `bump_event_types`. This helps clients sort more readily without them needing to
+        pull in a bunch of the timeline to determine the last activity.
+        `bump_event_types` is a thing because for example, we don't want display name
+        changes to mark the room as unread and bump it to the top. For encrypted rooms,
+        we just have to consider any activity as a bump because we can't see the content
+        and the client has to figure it out for themselves.
+
+        This should only be called where the server is participating
+        in the room (someone local is joined).
+
+        Returns:
+            The `bump_stamp` for the room (which can be `None`).
+        """
+
+        return cast(
+            Optional[int],
+            await self.db_pool.simple_select_one_onecol(
+                table="sliding_sync_joined_rooms",
+                keyvalues={"room_id": room_id},
+                retcol="bump_stamp",
+                # FIXME: This should be `False` once we bump `SCHEMA_COMPAT_VERSION` and run the
+                # foreground update for
+                # `sliding_sync_joined_rooms`/`sliding_sync_membership_snapshots` (tracked
+                # by https://github.com/element-hq/synapse/issues/17623)
+                #
+                # The should be `allow_none=False` in the future because event though
+                # `bump_stamp` itself can be `None`, we should have a row in the
+                # `sliding_sync_joined_rooms` table for any joined room.
+                allow_none=True,
+            ),
+        )
+
     async def persist_per_connection_state(
         self,
         user_id: str,