Backfill remote event fetched by MSC3030 so we can paginate from it later

changelog.d/13205.feature

@@ -0,0 +1 @@
+Allow pagination from remote event after discovering it from MSC3030 `/timestamp_to_event`.

synapse/federation/federation_client.py

@@ -53,7 +53,7 @@
     RoomVersion,
     RoomVersions,
 )
-from synapse.events import EventBase, builder
+from synapse.events import EventBase, builder, make_event_from_dict
 from synapse.federation.federation_base import (
     FederationBase,
     InvalidEventSignatureError,
@@ -309,7 +309,7 @@ async def get_pdu_from_destination_raw(
         )
 
         logger.debug(
-            "retrieved event id %s from %s: %r",
+            "get_pdu_from_destination_raw: retrieved event id %s from %s: %r",
             event_id,
             destination,
             transaction_data,
@@ -360,9 +360,25 @@ async def get_pdu(
 
         # TODO: Rate limit the number of times we try and get the same event.
 
-        ev = self._get_pdu_cache.get(event_id)
-        if ev:
-            return ev
+        event_from_cache = self._get_pdu_cache.get(event_id)
+        if event_from_cache:
+            assert not event_from_cache.internal_metadata.outlier, (
+                "Event from cache unexpectedly an `outlier` when it should be pristine and untouched without metadata set. "
+                "We are probably not be returning a copy of the event because downstream callers are modifying the event reference we have in the cache."
+            )
+
+            # Make sure to return a copy because downstream callers will use
+            # this event reference directly and change our original, pristine,
+            # untouched PDU. For example when people mark the event as an
+            # `outlier` (`event.internal_metadata.outlier = true`), we don't
+            # want that to propagate back into the cache.
+            event_copy = make_event_from_dict(
+                event_from_cache.get_pdu_json(),
+                event_from_cache.room_version,
+                internal_metadata_dict=None,
+            )
+
+            return event_copy
 
         pdu_attempts = self.pdu_destination_tried.setdefault(event_id, {})
 
@@ -405,7 +421,22 @@ async def get_pdu(
         if signed_pdu:
             self._get_pdu_cache[event_id] = signed_pdu
 
-        return signed_pdu
+            # Make sure to return a copy because downstream callers will use this
+            # event reference directly and change our original, pristine, untouched
+            # PDU. For example when people mark the event as an `outlier`
+            # (`event.internal_metadata.outlier = true`), we don't want that to
+            # propagate back into the cache.
+            #
+            # We could get away with only making a new copy of the event when
+            # pulling from cache but it's probably better to have good hygiene and
+            # not dirty the cache in the first place as well.
+            event_copy = make_event_from_dict(
+                signed_pdu.get_pdu_json(),
+                signed_pdu.room_version,
+                internal_metadata_dict=None,
+            )
+
+        return event_copy
 
     async def get_room_state_ids(
         self, destination: str, room_id: str, event_id: str

synapse/handlers/federation_event.py

@@ -766,10 +766,22 @@ async def _process_pulled_event(
         """
         logger.info("Processing pulled event %s", event)
 
-        # these should not be outliers.
-        assert (
-            not event.internal_metadata.is_outlier()
-        ), "pulled event unexpectedly flagged as outlier"
+        # This function should only be used to backfill events. If you're trying
+        # to persist an outlier, use another method. If you happen to run into a
+        # situation where the event you're trying to backfill is marked as an
+        # `outlier`, then you should update that spot to return an `EventBase`
+        # without the `outlier` flag set.
+        #
+        # `EventBase` is used to represent both an event we have not yet
+        # persisted, and one that we have persisted and now keep in the cache.
+        # In an ideal world this method would only be called with the first type
+        # of event, but it turns out that's not actually the case and for
+        # example, you could get an event from cache that is marked as an
+        # `outlier` (fix up that spot though).
+        assert not event.internal_metadata.is_outlier(), (
+            "This is a safe-guard to make sure you're not trying to persist an outlier using this function (use something else). "
+            "If you're trying to backfill an event, this is the right method but you need pass in an event copy that doesn't have `event.internal_metada.outlier = true`."
+        )
 
         event_id = event.event_id
 
@@ -779,7 +791,7 @@ async def _process_pulled_event(
         if existing:
             if not existing.internal_metadata.is_outlier():
                 logger.info(
-                    "Ignoring received event %s which we have already seen",
+                    "_process_pulled_event: Ignoring received event %s which we have already seen",
                     event_id,
                 )
                 return
@@ -1312,6 +1324,53 @@ async def _handle_marker_event(self, origin: str, marker_event: EventBase) -> No
             marker_event,
         )
 
+    async def backfill_event_id(
+        self, destination: str, room_id: str, event_id: str
+    ) -> EventBase:
+        """Backfill a single event and persist it as a non-outlier which means
+        we also pull in all of the state and auth events necessary for it.
+
+        Args:
+            destination: The homeserver to pull the given event_id from.
+            room_id: The room where the event is from.
+            event_id: The event ID to backfill.
+
+        Raises:
+            FederationError if we are unable to find the event from the destination
+        """
+        logger.info(
+            "backfill_event_id: event_id=%s from destination=%s", event_id, destination
+        )
+
+        room_version = await self._store.get_room_version(room_id)
+
+        event_from_response = await self._federation_client.get_pdu(
+            [destination],
+            event_id,
+            room_version,
+        )
+
+        if not event_from_response:
+            raise FederationError(
+                "ERROR",
+                404,
+                "Unable to find event_id=%s from destination=%s to backfill."
+                % (event_id, destination),
+                affected=event_id,
+            )
+
+        # Persist the event we just fetched, including pulling all of the state
+        # and auth events to de-outlier it. This also sets up the necessary
+        # `state_groups` for the event.
+        await self._process_pulled_events(
+            destination,
+            [event_from_response],
+            # Prevent notifications going to clients
+            backfilled=True,
+        )
+
+        return event_from_response
+
     async def _get_events_and_persist(
         self, destination: str, room_id: str, event_ids: Collection[str]
     ) -> None:

synapse/handlers/room.py

@@ -1384,6 +1384,7 @@ def __init__(self, hs: "HomeServer"):
         self.store = hs.get_datastores().main
         self.state_handler = hs.get_state_handler()
         self.federation_client = hs.get_federation_client()
+        self.federation_event_handler = hs.get_federation_event_handler()
         self._storage_controllers = hs.get_storage_controllers()
 
     async def get_event_for_timestamp(
@@ -1479,38 +1480,65 @@ async def get_event_for_timestamp(
                         remote_response,
                     )
 
-                    # TODO: Do we want to persist this as an extremity?
-                    # TODO: I think ideally, we would try to backfill from
-                    # this event and run this whole
-                    # `get_event_for_timestamp` function again to make sure
-                    # they didn't give us an event from their gappy history.
                     remote_event_id = remote_response.event_id
-                    origin_server_ts = remote_response.origin_server_ts
+                    remote_origin_server_ts = remote_response.origin_server_ts
+
+                    # Backfill this event so we can get a pagination token for
+                    # it with `/context` and paginate `/messages` from this
+                    # point.
+                    #
+                    # FIXME: After this backfill, we might want to run this
+                    # whole `get_event_for_timestamp` function again to make
+                    # sure they didn't give us an event from their gappy
+                    # history. Also need a heuristic for when to stop recursing
+                    # if they keep giving us gappy results.
+                    remote_event = (
+                        await self.federation_event_handler.backfill_event_id(
+                            domain, room_id, remote_event_id
+                        )
+                    )
+
+                    # XXX: When we see that the remote server is not trustworthy,
+                    # maybe we should not ask them first in the future.
+                    if remote_origin_server_ts != remote_event.origin_server_ts:
+                        logger.info(
+                            "get_event_for_timestamp: Remote server (%s) claimed that remote_event_id=%s occured at remote_origin_server_ts=%s but that isn't true (actually occured at %s). Their claims are dubious and we should consider not trusting them.",
+                            domain,
+                            remote_event_id,
+                            remote_origin_server_ts,
+                            remote_event.origin_server_ts,
+                        )
 
                     # Only return the remote event if it's closer than the local event
                     if not local_event or (
-                        abs(origin_server_ts - timestamp)
+                        abs(remote_event.origin_server_ts - timestamp)
                         < abs(local_event.origin_server_ts - timestamp)
                     ):
-                        return remote_event_id, origin_server_ts
+                        logger.info(
+                            "get_event_for_timestamp: returning remote_event_id=%s (%s) since it's closer to timestamp=%s than local_event=%s (%s)",
+                            remote_event_id,
+                            remote_event.origin_server_ts,
+                            timestamp,
+                            local_event.event_id if local_event else None,
+                            local_event.origin_server_ts if local_event else None,
+                        )
+                        return remote_event_id, remote_origin_server_ts
                 except (HttpResponseException, InvalidResponseError) as ex:
                     # Let's not put a high priority on some other homeserver
                     # failing to respond or giving a random response
                     logger.debug(
-                        "Failed to fetch /timestamp_to_event from %s because of exception(%s) %s args=%s",
+                        "get_event_for_timestamp: Failed to fetch /timestamp_to_event from %s because of exception(%s) %s args=%s",
                         domain,
                         type(ex).__name__,
                         ex,
                         ex.args,
                     )
-                except Exception as ex:
+                except Exception:
                     # But we do want to see some exceptions in our code
                     logger.warning(
-                        "Failed to fetch /timestamp_to_event from %s because of exception(%s) %s args=%s",
+                        "get_event_for_timestamp: Failed to fetch /timestamp_to_event from %s because of exception",
                         domain,
-                        type(ex).__name__,
-                        ex,
-                        ex.args,
+                        exc_info=True,
                     )
 
         # To appease mypy, we have to add both of these conditions to check for

synapse/storage/databases/main/events.py

@@ -1346,9 +1346,24 @@ def _update_outliers_txn(
             event_id: outlier for event_id, outlier in txn
         }
 
+        logger.debug(
+            "_update_outliers_txn: events=%s have_persisted=%s",
+            [ev.event_id for ev, _ in events_and_contexts],
+            have_persisted,
+        )
+
         to_remove = set()
         for event, context in events_and_contexts:
-            if event.event_id not in have_persisted:
+            outlier_persisted = have_persisted.get(event.event_id)
+            logger.debug(
+                "_update_outliers_txn: event=%s outlier=%s outlier_persisted=%s",
+                event.event_id,
+                event.internal_metadata.is_outlier(),
+                outlier_persisted,
+            )
+
+            # Ignore events which we haven't persisted at all
+            if outlier_persisted is None:
                 continue
 
             to_remove.add(event)
@@ -1358,7 +1373,6 @@ def _update_outliers_txn(
                 # was an outlier or not - what we have is at least as good.
                 continue
 
-            outlier_persisted = have_persisted[event.event_id]
             if not event.internal_metadata.is_outlier() and outlier_persisted:
                 # We received a copy of an event that we had already stored as
                 # an outlier in the database. We now have some state at that event
@@ -1369,7 +1383,10 @@ def _update_outliers_txn(
                 # events down /sync. In general they will be historical events, so that
                 # doesn't matter too much, but that is not always the case.
 
-                logger.info("Updating state for ex-outlier event %s", event.event_id)
+                logger.info(
+                    "_update_outliers_txn: Updating state for ex-outlier event %s",
+                    event.event_id,
+                )
 
                 # insert into event_to_state_groups.
                 try: