Add experimental support for sharding event persister. (#8170)

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.

changelog.d/8170.feature

@@ -0,0 +1 @@
+Add experimental support for sharding event persister.

synapse/config/_base.py

@@ -832,11 +832,26 @@ class ShardedWorkerHandlingConfig:
     def should_handle(self, instance_name: str, key: str) -> bool:
         """Whether this instance is responsible for handling the given key.
         """
-
-        # If multiple instances are not defined we always return true.
+        # If multiple instances are not defined we always return true
         if not self.instances or len(self.instances) == 1:
             return True
 
+        return self.get_instance(key) == instance_name
+
+    def get_instance(self, key: str) -> str:
+        """Get the instance responsible for handling the given key.
+
+        Note: For things like federation sending the config for which instance
+        is sending is known only to the sender instance if there is only one.
+        Therefore `should_handle` should be used where possible.
+        """
+
+        if not self.instances:
+            return "master"
+
+        if len(self.instances) == 1:
+            return self.instances[0]
+
         # We shard by taking the hash, modulo it by the number of instances and
         # then checking whether this instance matches the instance at that
         # index.
@@ -846,7 +861,7 @@ class ShardedWorkerHandlingConfig:
         dest_hash = sha256(key.encode("utf8")).digest()
         dest_int = int.from_bytes(dest_hash, byteorder="little")
         remainder = dest_int % (len(self.instances))
-        return self.instances[remainder] == instance_name
+        return self.instances[remainder]
 
 
 __all__ = ["Config", "RootConfig", "ShardedWorkerHandlingConfig"]

synapse/config/_base.pyi

@@ -142,3 +142,4 @@ class ShardedWorkerHandlingConfig:
     instances: List[str]
     def __init__(self, instances: List[str]) -> None: ...
     def should_handle(self, instance_name: str, key: str) -> bool: ...
+    def get_instance(self, key: str) -> str: ...

synapse/config/workers.py

@@ -13,12 +13,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import List, Union
+
 import attr
 
 from ._base import Config, ConfigError, ShardedWorkerHandlingConfig
 from .server import ListenerConfig, parse_listener_def
 
 
+def _instance_to_list_converter(obj: Union[str, List[str]]) -> List[str]:
+    """Helper for allowing parsing a string or list of strings to a config
+    option expecting a list of strings.
+    """
+
+    if isinstance(obj, str):
+        return [obj]
+    return obj
+
+
 @attr.s
 class InstanceLocationConfig:
     """The host and port to talk to an instance via HTTP replication.
@@ -33,11 +45,13 @@ class WriterLocations:
     """Specifies the instances that write various streams.
 
     Attributes:
-        events: The instance that writes to the event and backfill streams.
-        events: The instance that writes to the typing stream.
+        events: The instances that write to the event and backfill streams.
+        typing: The instance that writes to the typing stream.
     """
 
-    events = attr.ib(default="master", type=str)
+    events = attr.ib(
+        default=["master"], type=List[str], converter=_instance_to_list_converter
+    )
     typing = attr.ib(default="master", type=str)
 
 
@@ -105,16 +119,19 @@ class WorkerConfig(Config):
         writers = config.get("stream_writers") or {}
         self.writers = WriterLocations(**writers)
 
-        # Check that the configured writer for events and typing also appears in
+        # Check that the configured writers for events and typing also appears in
         # `instance_map`.
         for stream in ("events", "typing"):
-            instance = getattr(self.writers, stream)
+            instances = _instance_to_list_converter(getattr(self.writers, stream))
+            for instance in instances:
                 if instance != "master" and instance not in self.instance_map:
                     raise ConfigError(
                         "Instance %r is configured to write %s but does not appear in `instance_map` config."
                         % (instance, stream)
                     )
 
+        self.events_shard_config = ShardedWorkerHandlingConfig(self.writers.events)
+
     def generate_config_section(self, config_dir_path, server_name, **kwargs):
         return """\
         ## Workers ##

synapse/handlers/federation.py

@@ -923,7 +923,8 @@ 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)
@@ -1216,7 +1217,7 @@ 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):
@@ -1363,15 +1364,15 @@ 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
@@ -1625,7 +1626,7 @@ 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
 
@@ -1652,7 +1653,9 @@ 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
 
@@ -1900,7 +1903,7 @@ 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(
@@ -1913,6 +1916,7 @@ class FederationHandler(BaseHandler):
     async def _handle_new_events(
         self,
         origin: str,
+        room_id: str,
         event_infos: Iterable[_NewEventInfo],
         backfilled: bool = False,
     ) -> None:
@@ -1944,6 +1948,7 @@ class FederationHandler(BaseHandler):
         )
 
         await self.persist_events_and_notify(
+            room_id,
             [
                 (ev_info.event, context)
                 for ev_info, context in zip(event_infos, contexts)
@@ -1954,6 +1959,7 @@ class FederationHandler(BaseHandler):
     async def _persist_auth_tree(
         self,
         origin: str,
+        room_id: str,
         auth_events: List[EventBase],
         state: List[EventBase],
         event: EventBase,
@@ -1968,6 +1974,7 @@ class FederationHandler(BaseHandler):
 
         Args:
             origin: Where the events came from
+            room_id,
             auth_events
             state
             event
@@ -2042,17 +2049,20 @@ 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,
@@ -2903,6 +2913,7 @@ class FederationHandler(BaseHandler):
 
     async def persist_events_and_notify(
         self,
+        room_id: str,
         event_and_contexts: Sequence[Tuple[EventBase, EventContext]],
         backfilled: bool = False,
     ) -> int:
@@ -2910,14 +2921,19 @@ 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(object):
         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
 
@@ -904,9 +903,10 @@ class EventCreationHandler(object):
 
         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,
@@ -974,7 +974,9 @@ class EventCreationHandler(object):
 
         This should only be run on the instance in charge of persisting events.
         """
-        assert self._is_event_writer
+        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
@@ -1260,10 +1262,10 @@ class RoomShutdownHandler(object):
             # 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
@@ -1293,7 +1295,9 @@ class RoomShutdownHandler(object):
 
                 # 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(object):
         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,

synapse/replication/http/federation.py

@@ -65,10 +65,11 @@ class ReplicationFederationSendEventsRestServlet(ReplicationEndpoint):
         self.federation_handler = hs.get_handlers().federation_handler
 
     @staticmethod
-    async def _serialize_payload(store, event_and_contexts, backfilled):
+    async def _serialize_payload(store, room_id, event_and_contexts, backfilled):
         """
         Args:
             store
+            room_id (str)
             event_and_contexts (list[tuple[FrozenEvent, EventContext]])
             backfilled (bool): Whether or not the events are the result of
                 backfilling
@@ -88,7 +89,11 @@ class ReplicationFederationSendEventsRestServlet(ReplicationEndpoint):
                 }
             )
 
-        payload = {"events": event_payloads, "backfilled": backfilled}
+        payload = {
+            "events": event_payloads,
+            "backfilled": backfilled,
+            "room_id": room_id,
+        }
 
         return payload
 
@@ -96,6 +101,7 @@ class ReplicationFederationSendEventsRestServlet(ReplicationEndpoint):
         with Measure(self.clock, "repl_fed_send_events_parse"):
             content = parse_json_object_from_request(request)
 
+            room_id = content["room_id"]
             backfilled = content["backfilled"]
 
             event_payloads = content["events"]
@@ -120,7 +126,7 @@ class ReplicationFederationSendEventsRestServlet(ReplicationEndpoint):
         logger.info("Got %d events from federation", len(event_and_contexts))
 
         max_stream_id = await self.federation_handler.persist_events_and_notify(
-            event_and_contexts, backfilled
+            room_id, event_and_contexts, backfilled
         )
 
         return 200, {"max_stream_id": max_stream_id}

synapse/replication/tcp/handler.py

@@ -109,7 +109,7 @@ class ReplicationCommandHandler:
             if isinstance(stream, (EventsStream, BackfillStream)):
                 # Only add EventStream and BackfillStream as a source on the
                 # instance in charge of event persistence.
-                if hs.config.worker.writers.events == hs.get_instance_name():
+                if hs.get_instance_name() in hs.config.worker.writers.events:
                     self._streams_to_replicate.append(stream)
 
                 continue

synapse/replication/tcp/streams/events.py

@@ -19,7 +19,7 @@ from typing import List, Tuple, Type
 
 import attr
 
-from ._base import Stream, StreamUpdateResult, Token, current_token_without_instance
+from ._base import Stream, StreamUpdateResult, Token
 
 """Handling of the 'events' replication stream
 
@@ -117,7 +117,7 @@ class EventsStream(Stream):
         self._store = hs.get_datastore()
         super().__init__(
             hs.get_instance_name(),
-            current_token_without_instance(self._store.get_current_events_token),
+            self._store._stream_id_gen.get_current_token_for_writer,
             self._update_function,
         )
 

synapse/storage/databases/__init__.py

@@ -68,7 +68,7 @@ class Databases(object):
 
                     # If we're on a process that can persist events also
                     # instantiate a `PersistEventsStore`
-                    if hs.config.worker.writers.events == hs.get_instance_name():
+                    if hs.get_instance_name() in hs.config.worker.writers.events:
                         persist_events = PersistEventsStore(hs, database, main)
 
                 if "state" in database_config.databases:

synapse/storage/databases/main/event_federation.py

@@ -438,7 +438,7 @@ class EventFederationWorkerStore(EventsWorkerStore, SignatureWorkerStore, SQLBas
         """
 
         if stream_ordering <= self.stream_ordering_month_ago:
-            raise StoreError(400, "stream_ordering too old")
+            raise StoreError(400, "stream_ordering too old %s" % (stream_ordering,))
 
         sql = """
                 SELECT event_id FROM stream_ordering_to_exterm

synapse/storage/databases/main/events.py

@@ -97,6 +97,7 @@ class PersistEventsStore:
         self.store = main_data_store
         self.database_engine = db.engine
         self._clock = hs.get_clock()
+        self._instance_name = hs.get_instance_name()
 
         self._ephemeral_messages_enabled = hs.config.enable_ephemeral_messages
         self.is_mine_id = hs.is_mine_id
@@ -108,7 +109,7 @@ class PersistEventsStore:
 
         # This should only exist on instances that are configured to write
         assert (
-            hs.config.worker.writers.events == hs.get_instance_name()
+            hs.get_instance_name() in hs.config.worker.writers.events
         ), "Can only instantiate EventsStore on master"
 
     async def _persist_events_and_state_updates(
@@ -800,6 +801,7 @@ class PersistEventsStore:
             table="events",
             values=[
                 {
+                    "instance_name": self._instance_name,
                     "stream_ordering": event.internal_metadata.stream_ordering,
                     "topological_ordering": event.depth,
                     "depth": event.depth,

synapse/storage/databases/main/events_worker.py

@@ -42,7 +42,8 @@ from synapse.replication.tcp.streams import BackfillStream
 from synapse.replication.tcp.streams.events import EventsStream
 from synapse.storage._base import SQLBaseStore, db_to_json, make_in_list_sql_clause
 from synapse.storage.database import DatabasePool
-from synapse.storage.util.id_generators import StreamIdGenerator
+from synapse.storage.engines import PostgresEngine
+from synapse.storage.util.id_generators import MultiWriterIdGenerator, StreamIdGenerator
 from synapse.types import Collection, get_domain_from_id
 from synapse.util.caches.descriptors import Cache, cached
 from synapse.util.iterutils import batch_iter
@@ -78,9 +79,37 @@ class EventsWorkerStore(SQLBaseStore):
     def __init__(self, database: DatabasePool, db_conn, hs):
         super(EventsWorkerStore, self).__init__(database, db_conn, hs)
 
-        if hs.config.worker.writers.events == hs.get_instance_name():
-            # We are the process in charge of generating stream ids for events,
-            # so instantiate ID generators based on the database
+        if isinstance(database.engine, PostgresEngine):
+            # If we're using Postgres than we can use `MultiWriterIdGenerator`
+            # regardless of whether this process writes to the streams or not.
+            self._stream_id_gen = MultiWriterIdGenerator(
+                db_conn=db_conn,
+                db=database,
+                instance_name=hs.get_instance_name(),
+                table="events",
+                instance_column="instance_name",
+                id_column="stream_ordering",
+                sequence_name="events_stream_seq",
+            )
+            self._backfill_id_gen = MultiWriterIdGenerator(
+                db_conn=db_conn,
+                db=database,
+                instance_name=hs.get_instance_name(),
+                table="events",
+                instance_column="instance_name",
+                id_column="stream_ordering",
+                sequence_name="events_backfill_stream_seq",
+                positive=False,
+            )
+        else:
+            # We shouldn't be running in worker mode with SQLite, but its useful
+            # to support it for unit tests.
+            #
+            # If this process is the writer than we need to use
+            # `StreamIdGenerator`, otherwise we use `SlavedIdTracker` which gets
+            # updated over replication. (Multiple writers are not supported for
+            # SQLite).
+            if hs.get_instance_name() in hs.config.worker.writers.events:
                 self._stream_id_gen = StreamIdGenerator(
                     db_conn, "events", "stream_ordering",
                 )
@@ -92,10 +121,9 @@ class EventsWorkerStore(SQLBaseStore):
                     extra_tables=[("ex_outlier_stream", "event_stream_ordering")],
                 )
             else:
-            # Another process is in charge of persisting events and generating
-            # stream IDs: rely on the replication streams to let us know which
-            # IDs we can process.
-            self._stream_id_gen = SlavedIdTracker(db_conn, "events", "stream_ordering")
+                self._stream_id_gen = SlavedIdTracker(
+                    db_conn, "events", "stream_ordering"
+                )
                 self._backfill_id_gen = SlavedIdTracker(
                     db_conn, "events", "stream_ordering", step=-1
                 )

synapse/storage/databases/main/schema/delta/58/14events_instance_name.sql

@@ -0,0 +1,16 @@
+/* Copyright 2020 The Matrix.org Foundation C.I.C.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ALTER TABLE events ADD COLUMN instance_name TEXT;

synapse/storage/databases/main/schema/delta/58/14events_instance_name.sql.postgres

@@ -0,0 +1,26 @@
+/* Copyright 2020 The Matrix.org Foundation C.I.C.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+CREATE SEQUENCE IF NOT EXISTS events_stream_seq;
+
+SELECT setval('events_stream_seq', (
+    SELECT COALESCE(MAX(stream_ordering), 1) FROM events
+));
+
+CREATE SEQUENCE IF NOT EXISTS events_backfill_stream_seq;
+
+SELECT setval('events_backfill_stream_seq', (
+    SELECT COALESCE(-MIN(stream_ordering), 1) FROM events
+));

synapse/storage/util/id_generators.py

@@ -231,8 +231,12 @@ class MultiWriterIdGenerator:
         # gaps should be relatively rare it's still worth doing the book keeping
         # that allows us to skip forwards when there are gapless runs of
         # positions.
+        #
+        # We start at 1 here as a) the first generated stream ID will be 2, and
+        # b) other parts of the code assume that stream IDs are strictly greater
+        # than 0.
         self._persisted_upto_position = (
-            min(self._current_positions.values()) if self._current_positions else 0
+            min(self._current_positions.values()) if self._current_positions else 1
         )
         self._known_persisted_positions = []  # type: List[int]
 
@@ -362,9 +366,7 @@ class MultiWriterIdGenerator:
         equal to it have been successfully persisted.
         """
 
-        # Currently we don't support this operation, as it's not obvious how to
-        # condense the stream positions of multiple writers into a single int.
-        raise NotImplementedError()
+        return self.get_persisted_upto_position()
 
     def get_current_token_for_writer(self, instance_name: str) -> int:
         """Returns the position of the given writer.