Update get_pdu to return the original, pristine EventBase

changelog.d/13320.misc

@@ -0,0 +1 @@
+Fix `FederationClient.get_pdu()` returning events from the cache as `outliers` instead of original events we saw over federation.

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,
@@ -358,11 +358,33 @@ async def get_pdu(
             The requested PDU, or None if we were unable to find it.
         """
 
+        logger.debug(
+            "get_pdu: event_id=%s from destinations=%s", event_id, destinations
+        )
+
         # 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_copy = None
+        event_from_cache = self._get_pdu_cache.get(event_id)
+        if event_from_cache:
+            logger.debug("get_pdu: found event_from_cache=%s", 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, {})
 
@@ -371,6 +393,13 @@ async def get_pdu(
             now = self._clock.time_msec()
             last_attempt = pdu_attempts.get(destination, 0)
             if last_attempt + PDU_RETRY_TIME_MS > now:
+                logger.debug(
+                    "get_pdu: skipping destination=%s because we tried it recently last_attempt=%s and we only check every %s (now=%s)",
+                    destination,
+                    last_attempt,
+                    PDU_RETRY_TIME_MS,
+                    now,
+                )
                 continue
 
             try:
@@ -403,9 +432,24 @@ async def get_pdu(
                 continue
 
         if signed_pdu:
-            self._get_pdu_cache[event_id] = signed_pdu
+            self._get_pdu_cache[signed_pdu.event_id] = 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 signed_pdu
+        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 +766,24 @@ 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 not be used to persist outliers (use something
+        # else) because this does a bunch of operations that aren't necessary
+        # (extra work; in particular, it makes sure we have all the prev_events
+        # and resolves the state across those prev events). If you happen to run
+        # into a situation where the event you're trying to process/backfill is
+        # marked as an `outlier`, then you should update that spot to return an
+        # `EventBase` copy that doesn't have `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 we're not trying to persist an outlier using this function (use something else). "
+            "If you're trying to process/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
 

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:

tests/federation/test_federation_client.py

@@ -22,6 +22,7 @@
 from twisted.test.proto_helpers import MemoryReactor
 
 from synapse.api.room_versions import RoomVersions
+from synapse.events import EventBase
 from synapse.server import HomeServer
 from synapse.types import JsonDict
 from synapse.util import Clock
@@ -38,31 +39,35 @@ def prepare(self, reactor: MemoryReactor, clock: Clock, homeserver: HomeServer):
         self._mock_agent = mock.create_autospec(twisted.web.client.Agent, spec_set=True)
         homeserver.get_federation_http_client().agent = self._mock_agent
 
-    def test_get_room_state(self):
-        creator = f"@creator:{self.OTHER_SERVER_NAME}"
-        test_room_id = "!room_id"
+        # Move clock up to somewhat realistic time so the PDU destination retry
+        # works (`now` needs to be larger than `0 + PDU_RETRY_TIME_MS`).
+        self.reactor.advance(1000000000)
+
+        self.creator = f"@creator:{self.OTHER_SERVER_NAME}"
+        self.test_room_id = "!room_id"
 
+    def test_get_room_state(self):
         # mock up some events to use in the response.
         # In real life, these would have things in `prev_events` and `auth_events`, but that's
         # a bit annoying to mock up, and the code under test doesn't care, so we don't bother.
         create_event_dict = self.add_hashes_and_signatures_from_other_server(
             {
-                "room_id": test_room_id,
+                "room_id": self.test_room_id,
                 "type": "m.room.create",
                 "state_key": "",
-                "sender": creator,
-                "content": {"creator": creator},
+                "sender": self.creator,
+                "content": {"creator": self.creator},
                 "prev_events": [],
                 "auth_events": [],
                 "origin_server_ts": 500,
             }
         )
         member_event_dict = self.add_hashes_and_signatures_from_other_server(
             {
-                "room_id": test_room_id,
+                "room_id": self.test_room_id,
                 "type": "m.room.member",
-                "sender": creator,
-                "state_key": creator,
+                "sender": self.creator,
+                "state_key": self.creator,
                 "content": {"membership": "join"},
                 "prev_events": [],
                 "auth_events": [],
@@ -71,9 +76,9 @@ def test_get_room_state(self):
         )
         pl_event_dict = self.add_hashes_and_signatures_from_other_server(
             {
-                "room_id": test_room_id,
+                "room_id": self.test_room_id,
                 "type": "m.room.power_levels",
-                "sender": creator,
+                "sender": self.creator,
                 "state_key": "",
                 "content": {},
                 "prev_events": [],
@@ -103,7 +108,7 @@ def test_get_room_state(self):
         state_resp, auth_resp = self.get_success(
             self.hs.get_federation_client().get_room_state(
                 "yet_another_server",
-                test_room_id,
+                self.test_room_id,
                 "event_id",
                 RoomVersions.V9,
             )
@@ -130,6 +135,102 @@ def test_get_room_state(self):
             ["m.room.create", "m.room.member", "m.room.power_levels"],
         )
 
+    def test_get_pdu_returns_nothing_when_event_does_not_exist(self):
+        """No event should be returned when there the event does not exist"""
+        remote_pdu = self.get_success(
+            self.hs.get_federation_client().get_pdu(
+                ["yet_another_server"],
+                "event_should_not_exist",
+                RoomVersions.V9,
+            )
+        )
+        self.assertEqual(remote_pdu, None)
+
+    def test_get_pdu(self):
+        """Test to make sure an event is returned by `get_pdu()`"""
+        self._get_pdu_once()
+
+    def test_get_pdu_event_from_cache_is_pristine(self):
+        """Test that modifications made to events returned by `get_pdu()`
+        do not propagate back to to the internal cache (events returned should
+        be a copy).
+        """
+
+        # Get the PDU in the cache
+        remote_pdu = self._get_pdu_once()
+
+        # Modify the the event reference.
+        # This change should not make it back to the `_get_pdu_cache`.
+        remote_pdu.internal_metadata.outlier = True
+
+        # Get the event again. This time it should read it from cache.
+        remote_pdu2 = self.get_success(
+            self.hs.get_federation_client().get_pdu(
+                ["yet_another_server"],
+                remote_pdu.event_id,
+                RoomVersions.V9,
+            )
+        )
+
+        # Sanity check that we are working against the same event
+        self.assertEqual(remote_pdu.event_id, remote_pdu2.event_id)
+
+        # Make sure the event does not include modification from earlier
+        self.assertIsNotNone(remote_pdu2)
+        self.assertEqual(remote_pdu2.internal_metadata.outlier, False)
+
+    def _get_pdu_once(self) -> EventBase:
+        """Retrieve an event via `get_pdu()` and asserts that an event was returned.
+        Also used to prime the cache for subsequent test logic.
+        """
+        message_event_dict = self.add_hashes_and_signatures_from_other_server(
+            {
+                "room_id": self.test_room_id,
+                "type": "m.room.message",
+                "sender": self.creator,
+                "state_key": "",
+                "content": {},
+                "prev_events": [],
+                "auth_events": [],
+                "origin_server_ts": 700,
+                "depth": 10,
+            }
+        )
+
+        # mock up the response, and have the agent return it
+        self._mock_agent.request.side_effect = lambda *args, **kwargs: defer.succeed(
+            _mock_response(
+                {
+                    "origin": "yet_another_server",
+                    "origin_server_ts": 900,
+                    "pdus": [
+                        message_event_dict,
+                    ],
+                }
+            )
+        )
+
+        remote_pdu = self.get_success(
+            self.hs.get_federation_client().get_pdu(
+                ["yet_another_server"],
+                "event_id",
+                RoomVersions.V9,
+            )
+        )
+
+        # check the right call got made to the agent
+        self._mock_agent.request.assert_called_once_with(
+            b"GET",
+            b"matrix://yet_another_server/_matrix/federation/v1/event/event_id",
+            headers=mock.ANY,
+            bodyProducer=None,
+        )
+
+        self.assertIsNotNone(remote_pdu)
+        self.assertEqual(remote_pdu.internal_metadata.outlier, False)
+
+        return remote_pdu
+
 
 def _mock_response(resp: JsonDict):
     body = json.dumps(resp).encode("utf-8")