Update all stream IDs after processing replication rows (#14723)

This creates a new store method, `process_replication_position` that
is called after `process_replication_rows`. By moving stream ID advances
here this guarantees any relevant cache invalidations will have been
applied before the stream is advanced.

This avoids race conditions where Python switches between threads mid
way through processing the `process_replication_rows` method where stream
IDs may be advanced before caches are invalidated due to class resolution
ordering.

See this comment/issue for further discussion:
	#14158 (comment)

Author: Fizzadar

changelog.d/14723.bugfix

@@ -0,0 +1 @@
+Ensure stream IDs are always updated after caches get invalidated with workers. Contributed by Nick @ Beeper (@fizzadar).

synapse/replication/tcp/client.py

@@ -152,6 +152,9 @@ async def on_rdata(
             rows: a list of Stream.ROW_TYPE objects as returned by Stream.parse_row.
         """
         self.store.process_replication_rows(stream_name, instance_name, token, rows)
+        # NOTE: this must be called after process_replication_rows to ensure any
+        # cache invalidations are first handled before any stream ID advances.
+        self.store.process_replication_position(stream_name, instance_name, token)
 
         if self.send_handler:
             await self.send_handler.process_replication_rows(stream_name, token, rows)

synapse/storage/_base.py

@@ -57,7 +57,22 @@ def process_replication_rows(  # noqa: B027 (no-op by design)
         token: int,
         rows: Iterable[Any],
     ) -> None:
-        pass
+        """
+        Used by storage classes to invalidate caches based on incoming replication data. These
+        must not update any ID generators, use `process_replication_position`.
+        """
+
+    def process_replication_position(  # noqa: B027 (no-op by design)
+        self,
+        stream_name: str,
+        instance_name: str,
+        token: int,
+    ) -> None:
+        """
+        Used by storage classes to advance ID generators based on incoming replication data. This
+        is called after process_replication_rows such that caches are invalidated before any token
+        positions advance.
+        """
 
     def _invalidate_state_caches(
         self, room_id: str, members_changed: Collection[str]

synapse/storage/databases/main/account_data.py

@@ -436,10 +436,7 @@ def process_replication_rows(
         token: int,
         rows: Iterable[Any],
     ) -> None:
-        if stream_name == TagAccountDataStream.NAME:
-            self._account_data_id_gen.advance(instance_name, token)
-        elif stream_name == AccountDataStream.NAME:
-            self._account_data_id_gen.advance(instance_name, token)
+        if stream_name == AccountDataStream.NAME:
             for row in rows:
                 if not row.room_id:
                     self.get_global_account_data_by_type_for_user.invalidate(
@@ -454,6 +451,15 @@ def process_replication_rows(
 
         super().process_replication_rows(stream_name, instance_name, token, rows)
 
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == TagAccountDataStream.NAME:
+            self._account_data_id_gen.advance(instance_name, token)
+        elif stream_name == AccountDataStream.NAME:
+            self._account_data_id_gen.advance(instance_name, token)
+        super().process_replication_position(stream_name, instance_name, token)
+
     async def add_account_data_to_room(
         self, user_id: str, room_id: str, account_data_type: str, content: JsonDict
     ) -> int:

synapse/storage/databases/main/cache.py

@@ -164,9 +164,6 @@ def process_replication_rows(
                     backfilled=True,
                 )
         elif stream_name == CachesStream.NAME:
-            if self._cache_id_gen:
-                self._cache_id_gen.advance(instance_name, token)
-
             for row in rows:
                 if row.cache_func == CURRENT_STATE_CACHE_NAME:
                     if row.keys is None:
@@ -182,6 +179,14 @@ def process_replication_rows(
 
         super().process_replication_rows(stream_name, instance_name, token, rows)
 
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == CachesStream.NAME:
+            if self._cache_id_gen:
+                self._cache_id_gen.advance(instance_name, token)
+        super().process_replication_position(stream_name, instance_name, token)
+
     def _process_event_stream_row(self, token: int, row: EventsStreamRow) -> None:
         data = row.data
 

synapse/storage/databases/main/deviceinbox.py

@@ -157,6 +157,13 @@ def process_replication_rows(
                     )
         return super().process_replication_rows(stream_name, instance_name, token, rows)
 
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == ToDeviceStream.NAME:
+            self._device_inbox_id_gen.advance(instance_name, token)
+        super().process_replication_position(stream_name, instance_name, token)
+
     def get_to_device_stream_token(self) -> int:
         return self._device_inbox_id_gen.get_current_token()
 

synapse/storage/databases/main/devices.py

@@ -162,14 +162,21 @@ def process_replication_rows(
         self, stream_name: str, instance_name: str, token: int, rows: Iterable[Any]
     ) -> None:
         if stream_name == DeviceListsStream.NAME:
-            self._device_list_id_gen.advance(instance_name, token)
             self._invalidate_caches_for_devices(token, rows)
         elif stream_name == UserSignatureStream.NAME:
-            self._device_list_id_gen.advance(instance_name, token)
             for row in rows:
                 self._user_signature_stream_cache.entity_has_changed(row.user_id, token)
         return super().process_replication_rows(stream_name, instance_name, token, rows)
 
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == DeviceListsStream.NAME:
+            self._device_list_id_gen.advance(instance_name, token)
+        elif stream_name == UserSignatureStream.NAME:
+            self._device_list_id_gen.advance(instance_name, token)
+        super().process_replication_position(stream_name, instance_name, token)
+
     def _invalidate_caches_for_devices(
         self, token: int, rows: Iterable[DeviceListsStream.DeviceListsStreamRow]
     ) -> None:

synapse/storage/databases/main/events_worker.py

@@ -388,11 +388,7 @@ def process_replication_rows(
         token: int,
         rows: Iterable[Any],
     ) -> None:
-        if stream_name == EventsStream.NAME:
-            self._stream_id_gen.advance(instance_name, token)
-        elif stream_name == BackfillStream.NAME:
-            self._backfill_id_gen.advance(instance_name, -token)
-        elif stream_name == UnPartialStatedEventStream.NAME:
+        if stream_name == UnPartialStatedEventStream.NAME:
             for row in rows:
                 assert isinstance(row, UnPartialStatedEventStreamRow)
 
@@ -405,6 +401,15 @@ def process_replication_rows(
 
         super().process_replication_rows(stream_name, instance_name, token, rows)
 
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == EventsStream.NAME:
+            self._stream_id_gen.advance(instance_name, token)
+        elif stream_name == BackfillStream.NAME:
+            self._backfill_id_gen.advance(instance_name, -token)
+        super().process_replication_position(stream_name, instance_name, token)
+
     async def have_censored_event(self, event_id: str) -> bool:
         """Check if an event has been censored, i.e. if the content of the event has been erased
         from the database due to a redaction.

synapse/storage/databases/main/presence.py

@@ -439,8 +439,14 @@ def process_replication_rows(
         rows: Iterable[Any],
     ) -> None:
         if stream_name == PresenceStream.NAME:
-            self._presence_id_gen.advance(instance_name, token)
             for row in rows:
                 self.presence_stream_cache.entity_has_changed(row.user_id, token)
                 self._get_presence_for_user.invalidate((row.user_id,))
         return super().process_replication_rows(stream_name, instance_name, token, rows)
+
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == PresenceStream.NAME:
+            self._presence_id_gen.advance(instance_name, token)
+        super().process_replication_position(stream_name, instance_name, token)

synapse/storage/databases/main/push_rule.py

@@ -154,6 +154,13 @@ def process_replication_rows(
                 self.push_rules_stream_cache.entity_has_changed(row.user_id, token)
         return super().process_replication_rows(stream_name, instance_name, token, rows)
 
+    def process_replication_position(
+        self, stream_name: str, instance_name: str, token: int
+    ) -> None:
+        if stream_name == PushRulesStream.NAME:
+            self._push_rules_stream_id_gen.advance(instance_name, token)
+        super().process_replication_position(stream_name, instance_name, token)
+
     @cached(max_entries=5000)
     async def get_push_rules_for_user(self, user_id: str) -> FilteredPushRules:
         rows = await self.db_pool.simple_select_list(