Fix infinite loop in partial-state resync (#13353)

Make sure that we only pull out events from the db once they have no
prev-events with partial state.

changelog.d/13353.bugfix

@@ -0,0 +1 @@
+Fix a bug in the experimental faster-room-joins support which could cause it to get stuck in an infinite loop.

synapse/handlers/federation_event.py

@@ -569,15 +569,15 @@ class FederationEventHandler:
 
             if context is None or context.partial_state:
                 # this can happen if some or all of the event's prev_events still have
-                # partial state - ie, an event has an earlier stream_ordering than one
-                # or more of its prev_events, so we de-partial-state it before its
-                # prev_events.
+                # partial state. We were careful to only pick events from the db without
+                # partial-state prev events, so that implies that a prev event has
+                # been persisted (with partial state) since we did the query.
                 #
-                # TODO(faster_joins): we probably need to be more intelligent, and
-                #    exclude partial-state prev_events from consideration
-                #    https://github.com/matrix-org/synapse/issues/13001
+                # So, let's just ignore `event` for now; when we re-run the db query
+                # we should instead get its partial-state prev event, which we will
+                # de-partial-state, and then come back to event.
                 logger.warning(
-                    "%s still has partial state: can't de-partial-state it yet",
+                    "%s still has prev_events with partial state: can't de-partial-state it yet",
                     event.event_id,
                 )
                 return

synapse/storage/databases/main/events_worker.py

@@ -2110,11 +2110,29 @@ class EventsWorkerStore(SQLBaseStore):
     def _get_partial_state_events_batch_txn(
         txn: LoggingTransaction, room_id: str
     ) -> List[str]:
+        # we want to work through the events from oldest to newest, so
+        # we only want events whose prev_events do *not* have partial state - hence
+        # the 'NOT EXISTS' clause in the below.
+        #
+        # This is necessary because ordering by stream ordering isn't quite enough
+        # to ensure that we work from oldest to newest event (in particular,
+        # if an event is initially persisted as an outlier and later de-outliered,
+        # it can end up with a lower stream_ordering than its prev_events).
+        #
+        # Typically this means we'll only return one event per batch, but that's
+        # hard to do much about.
+        #
+        # See also: https://github.com/matrix-org/synapse/issues/13001
         txn.execute(
             """
             SELECT event_id FROM partial_state_events AS pse
                 JOIN events USING (event_id)
-            WHERE pse.room_id = ?
+            WHERE pse.room_id = ? AND
+               NOT EXISTS(
+                  SELECT 1 FROM event_edges AS ee
+                     JOIN partial_state_events AS prev_pse ON (prev_pse.event_id=ee.prev_event_id)
+                     WHERE ee.event_id=pse.event_id
+               )
             ORDER BY events.stream_ordering
             LIMIT 100
             """,