Fix limit logic for EventsStream

synapse/replication/tcp/handler.py

@@ -87,7 +87,9 @@ def __init__(self, hs):
             stream.NAME: stream(hs) for stream in STREAMS_MAP.values()
         }  # type: Dict[str, Stream]
 
-        self._position_linearizer = Linearizer("replication_position")
+        self._position_linearizer = Linearizer(
+            "replication_position", clock=self._clock
+        )
 
         # Map of stream to batched updates. See RdataCommand for info on how
         # batching works.

synapse/replication/tcp/streams/events.py

@@ -176,16 +176,30 @@ async def _update_function(
             from_token, upper_limit, target_row_count
         )  # type: List[Tuple]
 
-        # again, if we've hit the limit there, we'll need to limit the other sources
-        assert len(state_rows) < target_row_count
+        assert len(state_rows) <= target_row_count
+
+        # there can be more than one row per stream_id in that table, so if we hit
+        # the limit there, we'll need to truncate the results so that we have a complete
+        # set of changes for all the stream IDs we include.
         if len(state_rows) == target_row_count:
             assert state_rows[-1][0] <= upper_limit
-            upper_limit = state_rows[-1][0]
-            limited = True
+            upper_limit = state_rows[-1][0] - 1
+
+            # search for the point to truncate the list
+            for idx in range(len(state_rows) - 1, 0, -1):
+                if state_rows[idx - 1][0] <= upper_limit:
+                    state_rows = state_rows[:idx]
+                    break
+            else:
+                # bother. We didn't get a full set of changes for even a single
+                # stream id. let's run the query again, without a row limit, but for
+                # just one stream id.
+                upper_limit += 1
+                state_rows = await self._store.get_all_updated_current_state_deltas(
+                    from_token, upper_limit, limit=None
+                )
 
-            # FIXME: is it a given that there is only one row per stream_id in the
-            # state_deltas table (so that we can be sure that we have got all of the
-            # rows for upper_limit)?
+            limited = True
 
         # finally, fetch the ex-outliers rows. We assume there are few enough of these
         # not to bother with the limit.

synapse/storage/data_stores/main/events_worker.py

@@ -1084,15 +1084,23 @@ def get_all_new_backfill_event_rows(txn):
             "get_all_new_backfill_event_rows", get_all_new_backfill_event_rows
         )
 
-    def get_all_updated_current_state_deltas(self, from_token, to_token, limit):
+    def get_all_updated_current_state_deltas(
+        self, from_token: int, to_token: int, limit: Optional[int]
+    ):
         def get_all_updated_current_state_deltas_txn(txn):
             sql = """
                 SELECT stream_id, room_id, type, state_key, event_id
                 FROM current_state_delta_stream
                 WHERE ? < stream_id AND stream_id <= ?
-                ORDER BY stream_id ASC LIMIT ?
+                ORDER BY stream_id ASC
             """
-            txn.execute(sql, (from_token, to_token, limit))
+            params = [from_token, to_token]
+
+            if limit is not None:
+                sql += "LIMIT ?"
+                params.append(limit)
+
+            txn.execute(sql, params)
             return txn.fetchall()
 
         return self.db.runInteraction(