Add experimental support for sharding event persister. Again. (#8294)

This is *not* ready for production yet. Caveats:

1. We should write some tests...
2. The stream token that we use for events can get stalled at the minimum position of all writers. This means that new events may not be processed and e.g. sent down sync streams if a writer isn't writing or is slow.

synapse/handlers/federation.py

@@ -896,7 +896,8 @@ class FederationHandler(BaseHandler):
                 )
             )
 
-        await self._handle_new_events(dest, ev_infos, backfilled=True)
+        if ev_infos:
+            await self._handle_new_events(dest, room_id, ev_infos, backfilled=True)
 
         # Step 2: Persist the rest of the events in the chunk one by one
         events.sort(key=lambda e: e.depth)
@@ -1189,7 +1190,7 @@ class FederationHandler(BaseHandler):
             event_infos.append(_NewEventInfo(event, None, auth))
 
         await self._handle_new_events(
-            destination, event_infos,
+            destination, room_id, event_infos,
         )
 
     def _sanity_check_event(self, ev):
@@ -1336,15 +1337,15 @@ class FederationHandler(BaseHandler):
             )
 
             max_stream_id = await self._persist_auth_tree(
-                origin, auth_chain, state, event, room_version_obj
+                origin, room_id, auth_chain, state, event, room_version_obj
             )
 
             # We wait here until this instance has seen the events come down
             # replication (if we're using replication) as the below uses caches.
-            #
-            # TODO: Currently the events stream is written to from master
             await self._replication.wait_for_stream_position(
-                self.config.worker.writers.events, "events", max_stream_id
+                self.config.worker.events_shard_config.get_instance(room_id),
+                "events",
+                max_stream_id,
             )
 
             # Check whether this room is the result of an upgrade of a room we already know
@@ -1593,7 +1594,7 @@ class FederationHandler(BaseHandler):
         )
 
         context = await self.state_handler.compute_event_context(event)
-        await self.persist_events_and_notify([(event, context)])
+        await self.persist_events_and_notify(event.room_id, [(event, context)])
 
         return event
 
@@ -1620,7 +1621,9 @@ class FederationHandler(BaseHandler):
         await self.federation_client.send_leave(host_list, event)
 
         context = await self.state_handler.compute_event_context(event)
-        stream_id = await self.persist_events_and_notify([(event, context)])
+        stream_id = await self.persist_events_and_notify(
+            event.room_id, [(event, context)]
+        )
 
         return event, stream_id
 
@@ -1868,7 +1871,7 @@ class FederationHandler(BaseHandler):
                 )
 
             await self.persist_events_and_notify(
-                [(event, context)], backfilled=backfilled
+                event.room_id, [(event, context)], backfilled=backfilled
             )
         except Exception:
             run_in_background(
@@ -1881,6 +1884,7 @@ class FederationHandler(BaseHandler):
     async def _handle_new_events(
         self,
         origin: str,
+        room_id: str,
         event_infos: Iterable[_NewEventInfo],
         backfilled: bool = False,
     ) -> None:
@@ -1912,6 +1916,7 @@ class FederationHandler(BaseHandler):
         )
 
         await self.persist_events_and_notify(
+            room_id,
             [
                 (ev_info.event, context)
                 for ev_info, context in zip(event_infos, contexts)
@@ -1922,6 +1927,7 @@ class FederationHandler(BaseHandler):
     async def _persist_auth_tree(
         self,
         origin: str,
+        room_id: str,
         auth_events: List[EventBase],
         state: List[EventBase],
         event: EventBase,
@@ -1936,6 +1942,7 @@ class FederationHandler(BaseHandler):
 
         Args:
             origin: Where the events came from
+            room_id,
             auth_events
             state
             event
@@ -2010,17 +2017,20 @@ class FederationHandler(BaseHandler):
                 events_to_context[e.event_id].rejected = RejectedReason.AUTH_ERROR
 
         await self.persist_events_and_notify(
+            room_id,
             [
                 (e, events_to_context[e.event_id])
                 for e in itertools.chain(auth_events, state)
-            ]
+            ],
         )
 
         new_event_context = await self.state_handler.compute_event_context(
             event, old_state=state
         )
 
-        return await self.persist_events_and_notify([(event, new_event_context)])
+        return await self.persist_events_and_notify(
+            room_id, [(event, new_event_context)]
+        )
 
     async def _prep_event(
         self,
@@ -2871,6 +2881,7 @@ class FederationHandler(BaseHandler):
 
     async def persist_events_and_notify(
         self,
+        room_id: str,
         event_and_contexts: Sequence[Tuple[EventBase, EventContext]],
         backfilled: bool = False,
     ) -> int:
@@ -2878,14 +2889,19 @@ class FederationHandler(BaseHandler):
         necessary.
 
         Args:
-            event_and_contexts:
+            room_id: The room ID of events being persisted.
+            event_and_contexts: Sequence of events with their associated
+                context that should be persisted. All events must belong to
+                the same room.
             backfilled: Whether these events are a result of
                 backfilling or not
         """
-        if self.config.worker.writers.events != self._instance_name:
+        instance = self.config.worker.events_shard_config.get_instance(room_id)
+        if instance != self._instance_name:
             result = await self._send_events(
-                instance_name=self.config.worker.writers.events,
+                instance_name=instance,
                 store=self.store,
+                room_id=room_id,
                 event_and_contexts=event_and_contexts,
                 backfilled=backfilled,
             )

synapse/handlers/message.py

@@ -376,9 +376,8 @@ class EventCreationHandler:
         self.notifier = hs.get_notifier()
         self.config = hs.config
         self.require_membership_for_aliases = hs.config.require_membership_for_aliases
-        self._is_event_writer = (
-            self.config.worker.writers.events == hs.get_instance_name()
-        )
+        self._events_shard_config = self.config.worker.events_shard_config
+        self._instance_name = hs.get_instance_name()
 
         self.room_invite_state_types = self.hs.config.room_invite_state_types
 
@@ -902,9 +901,10 @@ class EventCreationHandler:
 
         try:
             # If we're a worker we need to hit out to the master.
-            if not self._is_event_writer:
+            writer_instance = self._events_shard_config.get_instance(event.room_id)
+            if writer_instance != self._instance_name:
                 result = await self.send_event(
-                    instance_name=self.config.worker.writers.events,
+                    instance_name=writer_instance,
                     event_id=event.event_id,
                     store=self.store,
                     requester=requester,
@@ -972,8 +972,10 @@ class EventCreationHandler:
 
         This should only be run on the instance in charge of persisting events.
         """
-        assert self._is_event_writer
         assert self.storage.persistence is not None
+        assert self._events_shard_config.should_handle(
+            self._instance_name, event.room_id
+        )
 
         if ratelimit:
             # We check if this is a room admin redacting an event so that we

synapse/handlers/room.py

@@ -804,7 +804,9 @@ class RoomCreationHandler(BaseHandler):
 
         # Always wait for room creation to progate before returning
         await self._replication.wait_for_stream_position(
-            self.hs.config.worker.writers.events, "events", last_stream_id
+            self.hs.config.worker.events_shard_config.get_instance(room_id),
+            "events",
+            last_stream_id,
         )
 
         return result, last_stream_id
@@ -1259,10 +1261,10 @@ class RoomShutdownHandler:
             # We now wait for the create room to come back in via replication so
             # that we can assume that all the joins/invites have propogated before
             # we try and auto join below.
-            #
-            # TODO: Currently the events stream is written to from master
             await self._replication.wait_for_stream_position(
-                self.hs.config.worker.writers.events, "events", stream_id
+                self.hs.config.worker.events_shard_config.get_instance(new_room_id),
+                "events",
+                stream_id,
             )
         else:
             new_room_id = None
@@ -1292,7 +1294,9 @@ class RoomShutdownHandler:
 
                 # Wait for leave to come in over replication before trying to forget.
                 await self._replication.wait_for_stream_position(
-                    self.hs.config.worker.writers.events, "events", stream_id
+                    self.hs.config.worker.events_shard_config.get_instance(room_id),
+                    "events",
+                    stream_id,
                 )
 
                 await self.room_member_handler.forget(target_requester.user, room_id)

synapse/handlers/room_member.py

@@ -82,13 +82,6 @@ class RoomMemberHandler:
         self._enable_lookup = hs.config.enable_3pid_lookup
         self.allow_per_room_profiles = self.config.allow_per_room_profiles
 
-        self._event_stream_writer_instance = hs.config.worker.writers.events
-        self._is_on_event_persistence_instance = (
-            self._event_stream_writer_instance == hs.get_instance_name()
-        )
-        if self._is_on_event_persistence_instance:
-            self.persist_event_storage = hs.get_storage().persistence
-
         self._join_rate_limiter_local = Ratelimiter(
             clock=self.clock,
             rate_hz=hs.config.ratelimiting.rc_joins_local.per_second,