Stop getting missing `prev_events` after we already know their signature is invalid (#13816)

While https://github.com/matrix-org/synapse/pull/13635 stops us from doing the slow thing after we've already done it once, this PR stops us from doing one of the slow things in the first place.

Related to
 - https://github.com/matrix-org/synapse/issues/13622
    - https://github.com/matrix-org/synapse/pull/13635
 - https://github.com/matrix-org/synapse/issues/13676

Part of https://github.com/matrix-org/synapse/issues/13356

Follow-up to https://github.com/matrix-org/synapse/pull/13815 which tracks event signature failures.

With this PR, we avoid the call to the costly `_get_state_ids_after_missing_prev_event` because the signature failure will count as an attempt before and we filter events based on the backoff before calling `_get_state_ids_after_missing_prev_event` now.

For example, this will save us 156s out of the 185s total that this `matrix.org` `/messages` request. If you want to see the full Jaeger trace of this, you can drag and drop this `trace.json` into your own Jaeger, https://gist.github.com/MadLittleMods/4b12d0d0afe88c2f65ffcc907306b761

To explain this exact scenario around `/messages` -> backfill, we call `/backfill` and first check the signatures of the 100 events. We see bad signature for `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` and `$zuOn2Rd2vsC7SUia3Hp3r6JSkSFKcc5j3QTTqW_0jDw` (both member events). Then we process the 98 events remaining that have valid signatures but one of the events references `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` as a `prev_event`. So we have to do the whole `_get_state_ids_after_missing_prev_event` rigmarole which pulls in those same events which fail again because the signatures are still invalid.

 - `backfill`
    - `outgoing-federation-request` `/backfill`
    - `_check_sigs_and_hash_and_fetch`
       - `_check_sigs_and_hash_and_fetch_one` for each event received over backfill
          - ❗ `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` fails with `Signature on retrieved event was invalid.`: `unable to verify signature for sender domain xxx: 401: Failed to find any key to satisfy: _FetchKeyRequest(...)`
          - ❗ `$zuOn2Rd2vsC7SUia3Hp3r6JSkSFKcc5j3QTTqW_0jDw` fails with `Signature on retrieved event was invalid.`: `unable to verify signature for sender domain xxx: 401: Failed to find any key to satisfy: _FetchKeyRequest(...)`
   - `_process_pulled_events`
      - `_process_pulled_event` for each validated event
         - ❗ Event `$Q0iMdqtz3IJYfZQU2Xk2WjB5NDF8Gg8cFSYYyKQgKJ0` references `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` as a `prev_event` which is missing so we try to get it
            - `_get_state_ids_after_missing_prev_event`
               - `outgoing-federation-request` `/state_ids`
               - ❗ `get_pdu` for `$luA4l7QHhf_jadH3mI-AyFqho0U2Q-IXXUbGSMq6h6M` which fails the signature check again
               - ❗ `get_pdu` for `$zuOn2Rd2vsC7SUia3Hp3r6JSkSFKcc5j3QTTqW_0jDw` which fails the signature check

7 files changed, 386 insertions, 2 deletions

diff --git a/changelog.d/13816.feature b/changelog.d/13816.feature
new file mode 100644
index 0000000000..5eaa936b08
--- /dev/null
+++ b/changelog.d/13816.feature
@@ -0,0 +1 @@
+Stop fetching missing `prev_events` after we already know their signature is invalid.
diff --git a/synapse/api/errors.py b/synapse/api/errors.py
index c606207569..e0873b1913 100644
--- a/synapse/api/errors.py
+++ b/synapse/api/errors.py
@@ -640,6 +640,27 @@ class FederationError(RuntimeError):
         }
 
 
+class FederationPullAttemptBackoffError(RuntimeError):
+    """
+    Raised to indicate that we are are deliberately not attempting to pull the given
+    event over federation because we've already done so recently and are backing off.
+
+    Attributes:
+        event_id: The event_id which we are refusing to pull
+        message: A custom error message that gives more context
+    """
+
+    def __init__(self, event_ids: List[str], message: Optional[str]):
+        self.event_ids = event_ids
+
+        if message:
+            error_message = message
+        else:
+            error_message = f"Not attempting to pull event_ids={self.event_ids} because we already tried to pull them recently (backing off)."
+
+        super().__init__(error_message)
+
+
 class HttpResponseException(CodeMessageException):
     """
     Represents an HTTP-level failure of an outbound request
diff --git a/synapse/handlers/federation.py b/synapse/handlers/federation.py
index 44e70c6c3c..5f7e0a1f79 100644
--- a/synapse/handlers/federation.py
+++ b/synapse/handlers/federation.py
@@ -45,6 +45,7 @@ from synapse.api.errors import (
     Codes,
     FederationDeniedError,
     FederationError,
+    FederationPullAttemptBackoffError,
     HttpResponseException,
     LimitExceededError,
     NotFoundError,
@@ -1720,7 +1721,22 @@ class FederationHandler:
                             destination, event
                         )
                         break
+                    except FederationPullAttemptBackoffError as exc:
+                        # Log a warning about why we failed to process the event (the error message
+                        # for `FederationPullAttemptBackoffError` is pretty good)
+                        logger.warning("_sync_partial_state_room: %s", exc)
+                        # We do not record a failed pull attempt when we backoff fetching a missing
+                        # `prev_event` because not being able to fetch the `prev_events` just means
+                        # we won't be able to de-outlier the pulled event. But we can still use an
+                        # `outlier` in the state/auth chain for another event. So we shouldn't stop
+                        # a downstream event from trying to pull it.
+                        #
+                        # This avoids a cascade of backoff for all events in the DAG downstream from
+                        # one event backoff upstream.
                     except FederationError as e:
+                        # TODO: We should `record_event_failed_pull_attempt` here,
+                        #   see https://github.com/matrix-org/synapse/issues/13700
+
                         if attempt == len(destinations) - 1:
                             # We have tried every remote server for this event. Give up.
                             # TODO(faster_joins) giving up isn't the right thing to do
diff --git a/synapse/handlers/federation_event.py b/synapse/handlers/federation_event.py
index f382961099..4300e8dd40 100644
--- a/synapse/handlers/federation_event.py
+++ b/synapse/handlers/federation_event.py
@@ -44,6 +44,7 @@ from synapse.api.errors import (
     AuthError,
     Codes,
     FederationError,
+    FederationPullAttemptBackoffError,
     HttpResponseException,
     RequestSendFailed,
     SynapseError,
@@ -567,6 +568,9 @@ class FederationEventHandler:
             event: partial-state event to be de-partial-stated
 
         Raises:
+            FederationPullAttemptBackoffError if we are are deliberately not attempting
+                to pull the given event over federation because we've already done so
+                recently and are backing off.
             FederationError if we fail to request state from the remote server.
         """
         logger.info("Updating state for %s", event.event_id)
@@ -901,6 +905,18 @@ class FederationEventHandler:
                     context,
                     backfilled=backfilled,
                 )
+        except FederationPullAttemptBackoffError as exc:
+            # Log a warning about why we failed to process the event (the error message
+            # for `FederationPullAttemptBackoffError` is pretty good)
+            logger.warning("_process_pulled_event: %s", exc)
+            # We do not record a failed pull attempt when we backoff fetching a missing
+            # `prev_event` because not being able to fetch the `prev_events` just means
+            # we won't be able to de-outlier the pulled event. But we can still use an
+            # `outlier` in the state/auth chain for another event. So we shouldn't stop
+            # a downstream event from trying to pull it.
+            #
+            # This avoids a cascade of backoff for all events in the DAG downstream from
+            # one event backoff upstream.
         except FederationError as e:
             await self._store.record_event_failed_pull_attempt(
                 event.room_id, event_id, str(e)
@@ -947,6 +963,9 @@ class FederationEventHandler:
             The event context.
 
         Raises:
+            FederationPullAttemptBackoffError if we are are deliberately not attempting
+                to pull the given event over federation because we've already done so
+                recently and are backing off.
             FederationError if we fail to get the state from the remote server after any
                 missing `prev_event`s.
         """
@@ -957,6 +976,18 @@ class FederationEventHandler:
         seen = await self._store.have_events_in_timeline(prevs)
         missing_prevs = prevs - seen
 
+        # If we've already recently attempted to pull this missing event, don't
+        # try it again so soon. Since we have to fetch all of the prev_events, we can
+        # bail early here if we find any to ignore.
+        prevs_to_ignore = await self._store.get_event_ids_to_not_pull_from_backoff(
+            room_id, missing_prevs
+        )
+        if len(prevs_to_ignore) > 0:
+            raise FederationPullAttemptBackoffError(
+                event_ids=prevs_to_ignore,
+                message=f"While computing context for event={event_id}, not attempting to pull missing prev_event={prevs_to_ignore[0]} because we already tried to pull recently (backing off).",
+            )
+
         if not missing_prevs:
             return await self._state_handler.compute_event_context(event)
 
diff --git a/synapse/storage/databases/main/event_federation.py b/synapse/storage/databases/main/event_federation.py
index 6b9a629edd..309a4ba664 100644
--- a/synapse/storage/databases/main/event_federation.py
+++ b/synapse/storage/databases/main/event_federation.py
@@ -1501,6 +1501,12 @@ class EventFederationWorkerStore(SignatureWorkerStore, EventsWorkerStore, SQLBas
             event_id: The event that failed to be fetched or processed
             cause: The error message or reason that we failed to pull the event
         """
+        logger.debug(
+            "record_event_failed_pull_attempt room_id=%s, event_id=%s, cause=%s",
+            room_id,
+            event_id,
+            cause,
+        )
         await self.db_pool.runInteraction(
             "record_event_failed_pull_attempt",
             self._record_event_failed_pull_attempt_upsert_txn,
@@ -1530,6 +1536,54 @@ class EventFederationWorkerStore(SignatureWorkerStore, EventsWorkerStore, SQLBas
 
         txn.execute(sql, (room_id, event_id, 1, self._clock.time_msec(), cause))
 
+    @trace
+    async def get_event_ids_to_not_pull_from_backoff(
+        self,
+        room_id: str,
+        event_ids: Collection[str],
+    ) -> List[str]:
+        """
+        Filter down the events to ones that we've failed to pull before recently. Uses
+        exponential backoff.
+
+        Args:
+            room_id: The room that the events belong to
+            event_ids: A list of events to filter down
+
+        Returns:
+            List of event_ids that should not be attempted to be pulled
+        """
+        event_failed_pull_attempts = await self.db_pool.simple_select_many_batch(
+            table="event_failed_pull_attempts",
+            column="event_id",
+            iterable=event_ids,
+            keyvalues={},
+            retcols=(
+                "event_id",
+                "last_attempt_ts",
+                "num_attempts",
+            ),
+            desc="get_event_ids_to_not_pull_from_backoff",
+        )
+
+        current_time = self._clock.time_msec()
+        return [
+            event_failed_pull_attempt["event_id"]
+            for event_failed_pull_attempt in event_failed_pull_attempts
+            # Exponential back-off (up to the upper bound) so we don't try to
+            # pull the same event over and over. ex. 2hr, 4hr, 8hr, 16hr, etc.
+            if current_time
+            < event_failed_pull_attempt["last_attempt_ts"]
+            + (
+                2
+                ** min(
+                    event_failed_pull_attempt["num_attempts"],
+                    BACKFILL_EVENT_EXPONENTIAL_BACKOFF_MAXIMUM_DOUBLING_STEPS,
+                )
+            )
+            * BACKFILL_EVENT_EXPONENTIAL_BACKOFF_STEP_MILLISECONDS
+        ]
+
     async def get_missing_events(
         self,
         room_id: str,
diff --git a/tests/handlers/test_federation_event.py b/tests/handlers/test_federation_event.py
index 918010cddb..e448cb1901 100644
--- a/tests/handlers/test_federation_event.py
+++ b/tests/handlers/test_federation_event.py
@@ -14,7 +14,7 @@
 from typing import Optional
 from unittest import mock
 
-from synapse.api.errors import AuthError
+from synapse.api.errors import AuthError, StoreError
 from synapse.api.room_versions import RoomVersion
 from synapse.event_auth import (
     check_state_dependent_auth_rules,
@@ -43,7 +43,7 @@ class FederationEventHandlerTests(unittest.FederatingHomeserverTestCase):
     def make_homeserver(self, reactor, clock):
         # mock out the federation transport client
         self.mock_federation_transport_client = mock.Mock(
-            spec=["get_room_state_ids", "get_room_state", "get_event"]
+            spec=["get_room_state_ids", "get_room_state", "get_event", "backfill"]
         )
         return super().setup_test_homeserver(
             federation_transport_client=self.mock_federation_transport_client
@@ -459,6 +459,203 @@ class FederationEventHandlerTests(unittest.FederatingHomeserverTestCase):
         )
         self.assertIsNotNone(persisted, "pulled event was not persisted at all")
 
+    def test_backfill_signature_failure_does_not_fetch_same_prev_event_later(
+        self,
+    ) -> None:
+        """
+        Test to make sure we backoff and don't try to fetch a missing prev_event when we
+        already know it has a invalid signature from checking the signatures of all of
+        the events in the backfill response.
+        """
+        OTHER_USER = f"@user:{self.OTHER_SERVER_NAME}"
+        main_store = self.hs.get_datastores().main
+
+        # Create the room
+        user_id = self.register_user("kermit", "test")
+        tok = self.login("kermit", "test")
+        room_id = self.helper.create_room_as(room_creator=user_id, tok=tok)
+        room_version = self.get_success(main_store.get_room_version(room_id))
+
+        # Allow the remote user to send state events
+        self.helper.send_state(
+            room_id,
+            "m.room.power_levels",
+            {"events_default": 0, "state_default": 0},
+            tok=tok,
+        )
+
+        # Add the remote user to the room
+        member_event = self.get_success(
+            event_injection.inject_member_event(self.hs, room_id, OTHER_USER, "join")
+        )
+
+        initial_state_map = self.get_success(
+            main_store.get_partial_current_state_ids(room_id)
+        )
+
+        auth_event_ids = [
+            initial_state_map[("m.room.create", "")],
+            initial_state_map[("m.room.power_levels", "")],
+            member_event.event_id,
+        ]
+
+        # We purposely don't run `add_hashes_and_signatures_from_other_server`
+        # over this because we want the signature check to fail.
+        pulled_event_without_signatures = make_event_from_dict(
+            {
+                "type": "test_regular_type",
+                "room_id": room_id,
+                "sender": OTHER_USER,
+                "prev_events": [member_event.event_id],
+                "auth_events": auth_event_ids,
+                "origin_server_ts": 1,
+                "depth": 12,
+                "content": {"body": "pulled_event_without_signatures"},
+            },
+            room_version,
+        )
+
+        # Create a regular event that should pass except for the
+        # `pulled_event_without_signatures` in the `prev_event`.
+        pulled_event = make_event_from_dict(
+            self.add_hashes_and_signatures_from_other_server(
+                {
+                    "type": "test_regular_type",
+                    "room_id": room_id,
+                    "sender": OTHER_USER,
+                    "prev_events": [
+                        member_event.event_id,
+                        pulled_event_without_signatures.event_id,
+                    ],
+                    "auth_events": auth_event_ids,
+                    "origin_server_ts": 1,
+                    "depth": 12,
+                    "content": {"body": "pulled_event"},
+                }
+            ),
+            room_version,
+        )
+
+        # We expect an outbound request to /backfill, so stub that out
+        self.mock_federation_transport_client.backfill.return_value = make_awaitable(
+            {
+                "origin": self.OTHER_SERVER_NAME,
+                "origin_server_ts": 123,
+                "pdus": [
+                    # This is one of the important aspects of this test: we include
+                    # `pulled_event_without_signatures` so it fails the signature check
+                    # when we filter down the backfill response down to events which
+                    # have valid signatures in
+                    # `_check_sigs_and_hash_for_pulled_events_and_fetch`
+                    pulled_event_without_signatures.get_pdu_json(),
+                    # Then later when we process this valid signature event, when we
+                    # fetch the missing `prev_event`s, we want to make sure that we
+                    # backoff and don't try and fetch `pulled_event_without_signatures`
+                    # again since we know it just had an invalid signature.
+                    pulled_event.get_pdu_json(),
+                ],
+            }
+        )
+
+        # Keep track of the count and make sure we don't make any of these requests
+        event_endpoint_requested_count = 0
+        room_state_ids_endpoint_requested_count = 0
+        room_state_endpoint_requested_count = 0
+
+        async def get_event(
+            destination: str, event_id: str, timeout: Optional[int] = None
+        ) -> None:
+            nonlocal event_endpoint_requested_count
+            event_endpoint_requested_count += 1
+
+        async def get_room_state_ids(
+            destination: str, room_id: str, event_id: str
+        ) -> None:
+            nonlocal room_state_ids_endpoint_requested_count
+            room_state_ids_endpoint_requested_count += 1
+
+        async def get_room_state(
+            room_version: RoomVersion, destination: str, room_id: str, event_id: str
+        ) -> None:
+            nonlocal room_state_endpoint_requested_count
+            room_state_endpoint_requested_count += 1
+
+        # We don't expect an outbound request to `/event`, `/state_ids`, or `/state` in
+        # the happy path but if the logic is sneaking around what we expect, stub that
+        # out so we can detect that failure
+        self.mock_federation_transport_client.get_event.side_effect = get_event
+        self.mock_federation_transport_client.get_room_state_ids.side_effect = (
+            get_room_state_ids
+        )
+        self.mock_federation_transport_client.get_room_state.side_effect = (
+            get_room_state
+        )
+
+        # The function under test: try to backfill and process the pulled event
+        with LoggingContext("test"):
+            self.get_success(
+                self.hs.get_federation_event_handler().backfill(
+                    self.OTHER_SERVER_NAME,
+                    room_id,
+                    limit=1,
+                    extremities=["$some_extremity"],
+                )
+            )
+
+        if event_endpoint_requested_count > 0:
+            self.fail(
+                "We don't expect an outbound request to /event in the happy path but if "
+                "the logic is sneaking around what we expect, make sure to fail the test. "
+                "We don't expect it because the signature failure should cause us to backoff "
+                "and not asking about pulled_event_without_signatures="
+                f"{pulled_event_without_signatures.event_id} again"
+            )
+
+        if room_state_ids_endpoint_requested_count > 0:
+            self.fail(
+                "We don't expect an outbound request to /state_ids in the happy path but if "
+                "the logic is sneaking around what we expect, make sure to fail the test. "
+                "We don't expect it because the signature failure should cause us to backoff "
+                "and not asking about pulled_event_without_signatures="
+                f"{pulled_event_without_signatures.event_id} again"
+            )
+
+        if room_state_endpoint_requested_count > 0:
+            self.fail(
+                "We don't expect an outbound request to /state in the happy path but if "
+                "the logic is sneaking around what we expect, make sure to fail the test. "
+                "We don't expect it because the signature failure should cause us to backoff "
+                "and not asking about pulled_event_without_signatures="
+                f"{pulled_event_without_signatures.event_id} again"
+            )
+
+        # Make sure we only recorded a single failure which corresponds to the signature
+        # failure initially in `_check_sigs_and_hash_for_pulled_events_and_fetch` before
+        # we process all of the pulled events.
+        backfill_num_attempts_for_event_without_signatures = self.get_success(
+            main_store.db_pool.simple_select_one_onecol(
+                table="event_failed_pull_attempts",
+                keyvalues={"event_id": pulled_event_without_signatures.event_id},
+                retcol="num_attempts",
+            )
+        )
+        self.assertEqual(backfill_num_attempts_for_event_without_signatures, 1)
+
+        # And make sure we didn't record a failure for the event that has the missing
+        # prev_event because we don't want to cause a cascade of failures. Not being
+        # able to fetch the `prev_events` just means we won't be able to de-outlier the
+        # pulled event. But we can still use an `outlier` in the state/auth chain for
+        # another event. So we shouldn't stop a downstream event from trying to pull it.
+        self.get_failure(
+            main_store.db_pool.simple_select_one_onecol(
+                table="event_failed_pull_attempts",
+                keyvalues={"event_id": pulled_event.event_id},
+                retcol="num_attempts",
+            ),
+            # StoreError: 404: No row found
+            StoreError,
+        )
+
     def test_process_pulled_event_with_rejected_missing_state(self) -> None:
         """Ensure that we correctly handle pulled events with missing state containing a
         rejected state event
diff --git a/tests/storage/test_event_federation.py b/tests/storage/test_event_federation.py
index 59b8910907..853db930d6 100644
--- a/tests/storage/test_event_federation.py
+++ b/tests/storage/test_event_federation.py
@@ -27,6 +27,8 @@ from synapse.api.room_versions import (
     RoomVersion,
 )
 from synapse.events import _EventInternalMetadata
+from synapse.rest import admin
+from synapse.rest.client import login, room
 from synapse.server import HomeServer
 from synapse.storage.database import LoggingTransaction
 from synapse.types import JsonDict
@@ -43,6 +45,12 @@ class _BackfillSetupInfo:
 
 
 class EventFederationWorkerStoreTestCase(tests.unittest.HomeserverTestCase):
+    servlets = [
+        admin.register_servlets,
+        room.register_servlets,
+        login.register_servlets,
+    ]
+
     def prepare(self, reactor: MemoryReactor, clock: Clock, hs: HomeServer) -> None:
         self.store = hs.get_datastores().main
 
@@ -1122,6 +1130,62 @@ class EventFederationWorkerStoreTestCase(tests.unittest.HomeserverTestCase):
         backfill_event_ids = [backfill_point[0] for backfill_point in backfill_points]
         self.assertEqual(backfill_event_ids, ["insertion_eventA"])
 
+    def test_get_event_ids_to_not_pull_from_backoff(
+        self,
+    ):
+        """
+        Test to make sure only event IDs we should backoff from are returned.
+        """
+        # Create the room
+        user_id = self.register_user("alice", "test")
+        tok = self.login("alice", "test")
+        room_id = self.helper.create_room_as(room_creator=user_id, tok=tok)
+
+        self.get_success(
+            self.store.record_event_failed_pull_attempt(
+                room_id, "$failed_event_id", "fake cause"
+            )
+        )
+
+        event_ids_to_backoff = self.get_success(
+            self.store.get_event_ids_to_not_pull_from_backoff(
+                room_id=room_id, event_ids=["$failed_event_id", "$normal_event_id"]
+            )
+        )
+
+        self.assertEqual(event_ids_to_backoff, ["$failed_event_id"])
+
+    def test_get_event_ids_to_not_pull_from_backoff_retry_after_backoff_duration(
+        self,
+    ):
+        """
+        Test to make sure no event IDs are returned after the backoff duration has
+        elapsed.
+        """
+        # Create the room
+        user_id = self.register_user("alice", "test")
+        tok = self.login("alice", "test")
+        room_id = self.helper.create_room_as(room_creator=user_id, tok=tok)
+
+        self.get_success(
+            self.store.record_event_failed_pull_attempt(
+                room_id, "$failed_event_id", "fake cause"
+            )
+        )
+
+        # Now advance time by 2 hours so we wait long enough for the single failed
+        # attempt (2^1 hours).
+        self.reactor.advance(datetime.timedelta(hours=2).total_seconds())
+
+        event_ids_to_backoff = self.get_success(
+            self.store.get_event_ids_to_not_pull_from_backoff(
+                room_id=room_id, event_ids=["$failed_event_id", "$normal_event_id"]
+            )
+        )
+        # Since this function only returns events we should backoff from, time has
+        # elapsed past the backoff range so there is no events to backoff from.
+        self.assertEqual(event_ids_to_backoff, [])
+
 
 @attr.s
 class FakeEvent: