Avoid checking the event cache when backfilling events (#14164)

2 files changed, 35 insertions, 14 deletions

diff --git a/synapse/handlers/federation_event.py b/synapse/handlers/federation_event.py
index 4300e8dd40..06e41b5cc0 100644
--- a/synapse/handlers/federation_event.py
+++ b/synapse/handlers/federation_event.py
@@ -798,9 +798,42 @@ class FederationEventHandler:
             ],
         )
 
+        # Check if we already any of these have these events.
+        # Note: we currently make a lookup in the database directly here rather than
+        # checking the event cache, due to:
+        # https://github.com/matrix-org/synapse/issues/13476
+        existing_events_map = await self._store._get_events_from_db(
+            [event.event_id for event in events]
+        )
+
+        new_events = []
+        for event in events:
+            event_id = event.event_id
+
+            # If we've already seen this event ID...
+            if event_id in existing_events_map:
+                existing_event = existing_events_map[event_id]
+
+                # ...and the event itself was not previously stored as an outlier...
+                if not existing_event.event.internal_metadata.is_outlier():
+                    # ...then there's no need to persist it. We have it already.
+                    logger.info(
+                        "_process_pulled_event: Ignoring received event %s which we "
+                        "have already seen",
+                        event.event_id,
+                    )
+                    continue
+
+                # While we have seen this event before, it was stored as an outlier.
+                # We'll now persist it as a non-outlier.
+                logger.info("De-outliering event %s", event_id)
+
+            # Continue on with the events that are new to us.
+            new_events.append(event)
+
         # We want to sort these by depth so we process them and
         # tell clients about them in order.
-        sorted_events = sorted(events, key=lambda x: x.depth)
+        sorted_events = sorted(new_events, key=lambda x: x.depth)
         for ev in sorted_events:
             with nested_logging_context(ev.event_id):
                 await self._process_pulled_event(origin, ev, backfilled=backfilled)
@@ -852,18 +885,6 @@ class FederationEventHandler:
 
         event_id = event.event_id
 
-        existing = await self._store.get_event(
-            event_id, allow_none=True, allow_rejected=True
-        )
-        if existing:
-            if not existing.internal_metadata.is_outlier():
-                logger.info(
-                    "_process_pulled_event: Ignoring received event %s which we have already seen",
-                    event_id,
-                )
-                return
-            logger.info("De-outliering event %s", event_id)
-
         try:
             self._sanity_check_event(event)
         except SynapseError as err:
diff --git a/synapse/storage/databases/main/events_worker.py b/synapse/storage/databases/main/events_worker.py
index cfd4780add..7bc7f2f33e 100644
--- a/synapse/storage/databases/main/events_worker.py
+++ b/synapse/storage/databases/main/events_worker.py
@@ -374,7 +374,7 @@ class EventsWorkerStore(SQLBaseStore):
                 If there is a mismatch, behave as per allow_none.
 
         Returns:
-            The event, or None if the event was not found.
+            The event, or None if the event was not found and allow_none is `True`.
         """
         if not isinstance(event_id, str):
             raise TypeError("Invalid event event_id %r" % (event_id,))