Fix races in room stats (and other) updates. (#6187)

Hopefully this will fix the occasional failures we were seeing in the room directory.

The problem was that events are not necessarily persisted (and `current_state_delta_stream` updated) in the same order as their stream_id. So for instance current_state_delta 9 might be persisted *before* current_state_delta 8. Then, when the room stats saw stream_id 9, it assumed it had done everything up to 9, and never came back to do stream_id 8.

We can solve this easily by only processing up to the stream_id where we know all events have been persisted.

4 files changed, 60 insertions, 23 deletions

diff --git a/synapse/handlers/presence.py b/synapse/handlers/presence.py
index 053cf66b28..2a5f1a007d 100644
--- a/synapse/handlers/presence.py
+++ b/synapse/handlers/presence.py
@@ -803,17 +803,25 @@ class PresenceHandler(object):
         # Loop round handling deltas until we're up to date
         while True:
             with Measure(self.clock, "presence_delta"):
-                deltas = yield self.store.get_current_state_deltas(self._event_pos)
-                if not deltas:
+                room_max_stream_ordering = self.store.get_room_max_stream_ordering()
+                if self._event_pos == room_max_stream_ordering:
                     return
 
+                logger.debug(
+                    "Processing presence stats %s->%s",
+                    self._event_pos,
+                    room_max_stream_ordering,
+                )
+                max_pos, deltas = yield self.store.get_current_state_deltas(
+                    self._event_pos, room_max_stream_ordering
+                )
                 yield self._handle_state_delta(deltas)
 
-                self._event_pos = deltas[-1]["stream_id"]
+                self._event_pos = max_pos
 
                 # Expose current event processing position to prometheus
                 synapse.metrics.event_processing_positions.labels("presence").set(
-                    self._event_pos
+                    max_pos
                 )
 
     @defer.inlineCallbacks
diff --git a/synapse/handlers/stats.py b/synapse/handlers/stats.py
index c62b113115..466daf9202 100644
--- a/synapse/handlers/stats.py
+++ b/synapse/handlers/stats.py
@@ -87,21 +87,23 @@ class StatsHandler(StateDeltasHandler):
             # Be sure to read the max stream_ordering *before* checking if there are any outstanding
             # deltas, since there is otherwise a chance that we could miss updates which arrive
             # after we check the deltas.
-            room_max_stream_ordering = yield self.store.get_room_max_stream_ordering()
+            room_max_stream_ordering = self.store.get_room_max_stream_ordering()
             if self.pos == room_max_stream_ordering:
                 break
 
-            deltas = yield self.store.get_current_state_deltas(self.pos)
+            logger.debug(
+                "Processing room stats %s->%s", self.pos, room_max_stream_ordering
+            )
+            max_pos, deltas = yield self.store.get_current_state_deltas(
+                self.pos, room_max_stream_ordering
+            )
 
             if deltas:
                 logger.debug("Handling %d state deltas", len(deltas))
                 room_deltas, user_deltas = yield self._handle_deltas(deltas)
-
-                max_pos = deltas[-1]["stream_id"]
             else:
                 room_deltas = {}
                 user_deltas = {}
-                max_pos = room_max_stream_ordering
 
             # Then count deltas for total_events and total_event_bytes.
             room_count, user_count = yield self.store.get_changes_room_total_events_and_bytes(
diff --git a/synapse/handlers/user_directory.py b/synapse/handlers/user_directory.py
index e53669e40d..624f05ab5b 100644
--- a/synapse/handlers/user_directory.py
+++ b/synapse/handlers/user_directory.py
@@ -138,21 +138,28 @@ class UserDirectoryHandler(StateDeltasHandler):
         # Loop round handling deltas until we're up to date
         while True:
             with Measure(self.clock, "user_dir_delta"):
-                deltas = yield self.store.get_current_state_deltas(self.pos)
-                if not deltas:
+                room_max_stream_ordering = self.store.get_room_max_stream_ordering()
+                if self.pos == room_max_stream_ordering:
                     return
 
+                logger.debug(
+                    "Processing user stats %s->%s", self.pos, room_max_stream_ordering
+                )
+                max_pos, deltas = yield self.store.get_current_state_deltas(
+                    self.pos, room_max_stream_ordering
+                )
+
                 logger.info("Handling %d state deltas", len(deltas))
                 yield self._handle_deltas(deltas)
 
-                self.pos = deltas[-1]["stream_id"]
+                self.pos = max_pos
 
                 # Expose current event processing position to prometheus
                 synapse.metrics.event_processing_positions.labels("user_dir").set(
-                    self.pos
+                    max_pos
                 )
 
-                yield self.store.update_user_directory_stream_pos(self.pos)
+                yield self.store.update_user_directory_stream_pos(max_pos)
 
     @defer.inlineCallbacks
     def _handle_deltas(self, deltas):
diff --git a/synapse/storage/state_deltas.py b/synapse/storage/state_deltas.py
index 5fdb442104..28f33ec18f 100644
--- a/synapse/storage/state_deltas.py
+++ b/synapse/storage/state_deltas.py
@@ -21,7 +21,7 @@ logger = logging.getLogger(__name__)
 
 
 class StateDeltasStore(SQLBaseStore):
-    def get_current_state_deltas(self, prev_stream_id):
+    def get_current_state_deltas(self, prev_stream_id: int, max_stream_id: int):
         """Fetch a list of room state changes since the given stream id
 
         Each entry in the result contains the following fields:
@@ -36,15 +36,27 @@ class StateDeltasStore(SQLBaseStore):
 
         Args:
             prev_stream_id (int): point to get changes since (exclusive)
+            max_stream_id (int): the point that we know has been correctly persisted
+               - ie, an upper limit to return changes from.
 
         Returns:
-            Deferred[list[dict]]: results
+            Deferred[tuple[int, list[dict]]: A tuple consisting of:
+               - the stream id which these results go up to
+               - list of current_state_delta_stream rows. If it is empty, we are
+                 up to date.
         """
         prev_stream_id = int(prev_stream_id)
+
+        # check we're not going backwards
+        assert prev_stream_id <= max_stream_id
+
         if not self._curr_state_delta_stream_cache.has_any_entity_changed(
             prev_stream_id
         ):
-            return []
+            # if the CSDs haven't changed between prev_stream_id and now, we
+            # know for certain that they haven't changed between prev_stream_id and
+            # max_stream_id.
+            return max_stream_id, []
 
         def get_current_state_deltas_txn(txn):
             # First we calculate the max stream id that will give us less than
@@ -54,21 +66,29 @@ class StateDeltasStore(SQLBaseStore):
             sql = """
                 SELECT stream_id, count(*)
                 FROM current_state_delta_stream
-                WHERE stream_id > ?
+                WHERE stream_id > ? AND stream_id <= ?
                 GROUP BY stream_id
                 ORDER BY stream_id ASC
                 LIMIT 100
             """
-            txn.execute(sql, (prev_stream_id,))
+            txn.execute(sql, (prev_stream_id, max_stream_id))
 
             total = 0
-            max_stream_id = prev_stream_id
-            for max_stream_id, count in txn:
+
+            for stream_id, count in txn:
                 total += count
                 if total > 100:
                     # We arbitarily limit to 100 entries to ensure we don't
                     # select toooo many.
+                    logger.debug(
+                        "Clipping current_state_delta_stream rows to stream_id %i",
+                        stream_id,
+                    )
+                    clipped_stream_id = stream_id
                     break
+            else:
+                # if there's no problem, we may as well go right up to the max_stream_id
+                clipped_stream_id = max_stream_id
 
             # Now actually get the deltas
             sql = """
@@ -77,8 +97,8 @@ class StateDeltasStore(SQLBaseStore):
                 WHERE ? < stream_id AND stream_id <= ?
                 ORDER BY stream_id ASC
             """
-            txn.execute(sql, (prev_stream_id, max_stream_id))
-            return self.cursor_to_dict(txn)
+            txn.execute(sql, (prev_stream_id, clipped_stream_id))
+            return clipped_stream_id, self.cursor_to_dict(txn)
 
         return self.runInteraction(
             "get_current_state_deltas", get_current_state_deltas_txn