Set # syntax=docker/dockerfile:experimental

For buildkit in some environments.

Signed-off-by: Jason Robinson <jasonr@matrix.org>

changelog.d/12522.misc

@@ -0,0 +1 @@
+Optimise room backfill, to reduce memory usage.

docker/Dockerfile

@@ -1,3 +1,4 @@
+# syntax=docker/dockerfile:experimental
 # Dockerfile to build the matrixdotorg/synapse docker images.
 #
 # Note that it uses features which are only available in BuildKit - see

synapse/handlers/federation.py

@@ -1,4 +1,4 @@
-# Copyright 2014-2021 The Matrix.org Foundation C.I.C.
+# Copyright 2014-2022 The Matrix.org Foundation C.I.C.
 # Copyright 2020 Sorunome
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -15,9 +15,10 @@
 
 """Contains handlers for federation events."""
 
+import itertools
 import logging
 from http import HTTPStatus
-from typing import TYPE_CHECKING, Dict, Iterable, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Dict, Iterable, List, Optional, Set, Tuple, Union
 
 from signedjson.key import decode_verify_key_bytes
 from signedjson.sign import verify_signed_json
@@ -178,68 +179,24 @@ class FederationHandler:
             logger.debug("Not backfilling as no extremeties found.")
             return False
 
-        # We only want to paginate if we can actually see the events we'll get,
-        # as otherwise we'll just spend a lot of resources to get redacted
-        # events.
-        #
-        # We do this by filtering all the backwards extremities and seeing if
-        # any remain. Given we don't have the extremity events themselves, we
-        # need to actually check the events that reference them.
-        #
-        # *Note*: the spec wants us to keep backfilling until we reach the start
-        # of the room in case we are allowed to see some of the history. However
-        # in practice that causes more issues than its worth, as a) its
-        # relatively rare for there to be any visible history and b) even when
-        # there is its often sufficiently long ago that clients would stop
-        # attempting to paginate before backfill reached the visible history.
-        #
-        # TODO: If we do do a backfill then we should filter the backwards
-        #   extremities to only include those that point to visible portions of
-        #   history.
-        #
-        # TODO: Correctly handle the case where we are allowed to see the
-        #   forward event but not the backward extremity, e.g. in the case of
-        #   initial join of the server where we are allowed to see the join
-        #   event but not anything before it. This would require looking at the
-        #   state *before* the event, ignoring the special casing certain event
-        #   types have.
-
-        forward_event_ids = await self.store.get_successor_events(
-            list(oldest_events_with_depth)
-        )
-
-        extremities_events = await self.store.get_events(
-            forward_event_ids,
-            redact_behaviour=EventRedactBehaviour.AS_IS,
-            get_prev_content=False,
-        )
-
-        # We set `check_history_visibility_only` as we might otherwise get false
-        # positives from users having been erased.
-        filtered_extremities = await filter_events_for_server(
-            self.storage,
-            self.server_name,
-            list(extremities_events.values()),
-            redact=False,
-            check_history_visibility_only=True,
+        # we now have a list of potential places to backpaginate from. We prefer to
+        # start with the most recent (ie, max depth), so let's sort the list.
+        sorted_extremeties_tuples: List[Tuple[str, int]] = sorted(
+            itertools.chain(
+                oldest_events_with_depth.items(),
+                insertion_events_to_be_backfilled.items(),
+            ),
+            key=lambda e: -int(e[1]),
         )
         logger.debug(
-            "_maybe_backfill_inner: filtered_extremities %s", filtered_extremities
+            "_maybe_backfill_inner: room_id: %s: current_depth: %s, limit: %s, extrems (%d): %s",
+            room_id,
+            current_depth,
+            limit,
+            len(sorted_extremeties_tuples),
+            sorted_extremeties_tuples,
         )
 
-        if not filtered_extremities and not insertion_events_to_be_backfilled:
-            return False
-
-        extremities = {
-            **oldest_events_with_depth,
-            # TODO: insertion_events_to_be_backfilled is currently skipping the filtered_extremities checks
-            **insertion_events_to_be_backfilled,
-        }
-
-        # Check if we reached a point where we should start backfilling.
-        sorted_extremeties_tuple = sorted(extremities.items(), key=lambda e: -int(e[1]))
-        max_depth = sorted_extremeties_tuple[0][1]
-
         # If we're approaching an extremity we trigger a backfill, otherwise we
         # no-op.
         #
@@ -249,6 +206,11 @@ class FederationHandler:
         # chose more than one times the limit in case of failure, but choosing a
         # much larger factor will result in triggering a backfill request much
         # earlier than necessary.
+        #
+        # XXX: shouldn't we do this *after* the filter by depth below? Again, we don't
+        # care about events that have happened after our current position.
+        #
+        max_depth = sorted_extremeties_tuples[0][1]
         if current_depth - 2 * limit > max_depth:
             logger.debug(
                 "Not backfilling as we don't need to. %d < %d - 2 * %d",
@@ -265,31 +227,98 @@ class FederationHandler:
         #    2. we have likely previously tried and failed to backfill from that
         #       extremity, so to avoid getting "stuck" requesting the same
         #       backfill repeatedly we drop those extremities.
-        filtered_sorted_extremeties_tuple = [
-            t for t in sorted_extremeties_tuple if int(t[1]) <= current_depth
-        ]
-
-        logger.debug(
-            "room_id: %s, backfill: current_depth: %s, limit: %s, max_depth: %s, extrems (%d): %s filtered_sorted_extremeties_tuple: %s",
-            room_id,
-            current_depth,
-            limit,
-            max_depth,
-            len(sorted_extremeties_tuple),
-            sorted_extremeties_tuple,
-            filtered_sorted_extremeties_tuple,
-        )
-
+        #
         # However, we need to check that the filtered extremities are non-empty.
         # If they are empty then either we can a) bail or b) still attempt to
         # backfill. We opt to try backfilling anyway just in case we do get
         # relevant events.
-        if filtered_sorted_extremeties_tuple:
-            sorted_extremeties_tuple = filtered_sorted_extremeties_tuple
+        #
+        filtered_sorted_extremeties_tuples = [
+            t for t in sorted_extremeties_tuples if int(t[1]) <= current_depth
+        ]
+        if filtered_sorted_extremeties_tuples:
+            logger.debug(
+                "_maybe_backfill_inner: extrems before current depth: %s",
+                filtered_sorted_extremeties_tuples,
+            )
+            sorted_extremeties_tuples = filtered_sorted_extremeties_tuples
+        else:
+            logger.debug(
+                "_maybe_backfill_inner: all extrems are *after* current depth. Backfilling anyway."
+            )
 
-        # We don't want to specify too many extremities as it causes the backfill
-        # request URI to be too long.
-        extremities = dict(sorted_extremeties_tuple[:5])
+        # We still need to narrow down the list of extremities we pass to the remote
+        # server. We limit to 5 of them, to avoid the request URI becoming too long.
+        #
+        # However, we only want to paginate from a particular extremity if we can
+        # actually see the events we'll get, as otherwise we'll just spend a lot of
+        # resources to get redacted events.
+        #
+        # We do this by filtering all the backwards extremities and seeing if
+        # any remain. Given we don't have the extremity events themselves, we
+        # need to actually check the events that reference them - their "successor"
+        # events.
+        #
+        # *Note*: the spec wants us to keep backfilling until we reach the start
+        # of the room in case we are allowed to see some of the history. However
+        # in practice that causes more issues than its worth, as a) its
+        # relatively rare for there to be any visible history and b) even when
+        # there is its often sufficiently long ago that clients would stop
+        # attempting to paginate before backfill reached the visible history.
+        #
+        # Calculating the visibility of each extremity is quite expensive, and there
+        # can be thousands of them in a big gappy room, so we just check them one
+        # by one until we've checked them all, or we've got 5 of them.
+        #
+        # TODO: Correctly handle the case where we are allowed to see the
+        #   successor event but not the backward extremity, e.g. in the case of
+        #   initial join of the server where we are allowed to see the join
+        #   event but not anything before it. This would require looking at the
+        #   state *before* the event, ignoring the special casing certain event
+        #   types have.
+
+        extremities_to_request: Set[str] = set()
+        for extremity_event_id, _ in sorted_extremeties_tuples:
+            if len(extremities_to_request) >= 5:
+                break
+
+            successor_event_ids = await self.store.get_successor_events(
+                [extremity_event_id]
+            )
+
+            successor_events = await self.store.get_events_as_list(
+                successor_event_ids,
+                redact_behaviour=EventRedactBehaviour.AS_IS,
+                get_prev_content=False,
+            )
+
+            # We set `check_history_visibility_only` as we might otherwise get false
+            # positives from users having been erased.
+            filtered_extremities = await filter_events_for_server(
+                self.storage,
+                self.server_name,
+                successor_events,
+                redact=False,
+                check_history_visibility_only=True,
+            )
+
+            if filtered_extremities:
+                extremities_to_request.add(extremity_event_id)
+            else:
+                logger.debug(
+                    "_maybe_backfill_inner: skipping extremity %s as it would not be visible.",
+                    extremity_event_id,
+                )
+
+        if not extremities_to_request:
+            logger.debug(
+                "_maybe_backfill_inner: found no extremities which would be visible"
+            )
+            return False
+
+        logger.debug(
+            "_maybe_backfill_inner: extremities_to_request %s", extremities_to_request
+        )
 
         # Now we need to decide which hosts to hit first.
 
@@ -309,7 +338,7 @@ class FederationHandler:
             for dom in domains:
                 try:
                     await self._federation_event_handler.backfill(
-                        dom, room_id, limit=100, extremities=extremities
+                        dom, room_id, limit=100, extremities=extremities_to_request
                     )
                     # If this succeeded then we probably already have the
                     # appropriate stuff.

synapse/visibility.py

@@ -419,6 +419,13 @@ async def _event_to_memberships(
         return {}
 
     # for each event, get the event_ids of the membership state at those events.
+    #
+    # TODO: this means that we request the entire membership list. If there  are only
+    #   one or two users on this server, and the room is huge, this is very wasteful
+    #   (it means more db work, and churns the *stateGroupMembersCache*).
+    #   It might be that we could extend StateFilter to specify "give me keys matching
+    #   *:<server_name>", to avoid this.
+
     event_to_state_ids = await storage.state.get_state_ids_for_events(
         frozenset(e.event_id for e in events),
         state_filter=StateFilter.from_types(types=((EventTypes.Member, None),)),