summary refs log tree commit diff
diff options
context:
space:
mode:
-rw-r--r--changelog.d/5278.bugfix1
-rw-r--r--synapse/storage/__init__.py2
-rw-r--r--synapse/storage/_base.py12
-rw-r--r--synapse/storage/events.py180
-rw-r--r--synapse/storage/events_bg_updates.py401
-rw-r--r--synapse/storage/schema/delta/54/delete_forward_extremities.sql22
-rw-r--r--tests/storage/test_cleanup_extrems.py248
7 files changed, 685 insertions, 181 deletions
diff --git a/changelog.d/5278.bugfix b/changelog.d/5278.bugfix
new file mode 100644
index 0000000000..9e14d20289
--- /dev/null
+++ b/changelog.d/5278.bugfix
@@ -0,0 +1 @@
+Fix bug where we leaked extremities when we soft failed events, leading to performance degradation.
diff --git a/synapse/storage/__init__.py b/synapse/storage/__init__.py
index 66675d08ae..71316f7d09 100644
--- a/synapse/storage/__init__.py
+++ b/synapse/storage/__init__.py
@@ -36,6 +36,7 @@ from .engines import PostgresEngine
 from .event_federation import EventFederationStore
 from .event_push_actions import EventPushActionsStore
 from .events import EventsStore
+from .events_bg_updates import EventsBackgroundUpdatesStore
 from .filtering import FilteringStore
 from .group_server import GroupServerStore
 from .keys import KeyStore
@@ -66,6 +67,7 @@ logger = logging.getLogger(__name__)
 
 
 class DataStore(
+    EventsBackgroundUpdatesStore,
     RoomMemberStore,
     RoomStore,
     RegistrationStore,
diff --git a/synapse/storage/_base.py b/synapse/storage/_base.py
index fa6839ceca..3fe827cd43 100644
--- a/synapse/storage/_base.py
+++ b/synapse/storage/_base.py
@@ -1261,7 +1261,8 @@ class SQLBaseStore(object):
             " AND ".join("%s = ?" % (k,) for k in keyvalues),
         )
 
-        return txn.execute(sql, list(keyvalues.values()))
+        txn.execute(sql, list(keyvalues.values()))
+        return txn.rowcount
 
     def _simple_delete_many(self, table, column, iterable, keyvalues, desc):
         return self.runInteraction(
@@ -1280,9 +1281,12 @@ class SQLBaseStore(object):
             column : column name to test for inclusion against `iterable`
             iterable : list
             keyvalues : dict of column names and values to select the rows with
+
+        Returns:
+            int: Number rows deleted
         """
         if not iterable:
-            return
+            return 0
 
         sql = "DELETE FROM %s" % table
 
@@ -1297,7 +1301,9 @@ class SQLBaseStore(object):
 
         if clauses:
             sql = "%s WHERE %s" % (sql, " AND ".join(clauses))
-        return txn.execute(sql, values)
+        txn.execute(sql, values)
+
+        return txn.rowcount
 
     def _get_cache_dict(
         self, db_conn, table, entity_column, stream_column, max_value, limit=100000
diff --git a/synapse/storage/events.py b/synapse/storage/events.py
index 6e9f3d1dc0..f9162be9b9 100644
--- a/synapse/storage/events.py
+++ b/synapse/storage/events.py
@@ -1,6 +1,7 @@
 # -*- coding: utf-8 -*-
 # Copyright 2014-2016 OpenMarket Ltd
-# Copyright 2018 New Vector Ltd
+# Copyright 2018-2019 New Vector Ltd
+# Copyright 2019 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.
@@ -219,41 +220,11 @@ class EventsStore(
     EventsWorkerStore,
     BackgroundUpdateStore,
 ):
-    EVENT_ORIGIN_SERVER_TS_NAME = "event_origin_server_ts"
-    EVENT_FIELDS_SENDER_URL_UPDATE_NAME = "event_fields_sender_url"
 
     def __init__(self, db_conn, hs):
         super(EventsStore, self).__init__(db_conn, hs)
-        self.register_background_update_handler(
-            self.EVENT_ORIGIN_SERVER_TS_NAME, self._background_reindex_origin_server_ts
-        )
-        self.register_background_update_handler(
-            self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME,
-            self._background_reindex_fields_sender,
-        )
-
-        self.register_background_index_update(
-            "event_contains_url_index",
-            index_name="event_contains_url_index",
-            table="events",
-            columns=["room_id", "topological_ordering", "stream_ordering"],
-            where_clause="contains_url = true AND outlier = false",
-        )
-
-        # an event_id index on event_search is useful for the purge_history
-        # api. Plus it means we get to enforce some integrity with a UNIQUE
-        # clause
-        self.register_background_index_update(
-            "event_search_event_id_idx",
-            index_name="event_search_event_id_idx",
-            table="event_search",
-            columns=["event_id"],
-            unique=True,
-            psql_only=True,
-        )
 
         self._event_persist_queue = _EventPeristenceQueue()
-
         self._state_resolution_handler = hs.get_state_resolution_handler()
 
     @defer.inlineCallbacks
@@ -1579,153 +1550,6 @@ class EventsStore(
         ret = yield self.runInteraction("count_daily_active_rooms", _count)
         defer.returnValue(ret)
 
-    @defer.inlineCallbacks
-    def _background_reindex_fields_sender(self, progress, batch_size):
-        target_min_stream_id = progress["target_min_stream_id_inclusive"]
-        max_stream_id = progress["max_stream_id_exclusive"]
-        rows_inserted = progress.get("rows_inserted", 0)
-
-        INSERT_CLUMP_SIZE = 1000
-
-        def reindex_txn(txn):
-            sql = (
-                "SELECT stream_ordering, event_id, json FROM events"
-                " INNER JOIN event_json USING (event_id)"
-                " WHERE ? <= stream_ordering AND stream_ordering < ?"
-                " ORDER BY stream_ordering DESC"
-                " LIMIT ?"
-            )
-
-            txn.execute(sql, (target_min_stream_id, max_stream_id, batch_size))
-
-            rows = txn.fetchall()
-            if not rows:
-                return 0
-
-            min_stream_id = rows[-1][0]
-
-            update_rows = []
-            for row in rows:
-                try:
-                    event_id = row[1]
-                    event_json = json.loads(row[2])
-                    sender = event_json["sender"]
-                    content = event_json["content"]
-
-                    contains_url = "url" in content
-                    if contains_url:
-                        contains_url &= isinstance(content["url"], text_type)
-                except (KeyError, AttributeError):
-                    # If the event is missing a necessary field then
-                    # skip over it.
-                    continue
-
-                update_rows.append((sender, contains_url, event_id))
-
-            sql = "UPDATE events SET sender = ?, contains_url = ? WHERE event_id = ?"
-
-            for index in range(0, len(update_rows), INSERT_CLUMP_SIZE):
-                clump = update_rows[index : index + INSERT_CLUMP_SIZE]
-                txn.executemany(sql, clump)
-
-            progress = {
-                "target_min_stream_id_inclusive": target_min_stream_id,
-                "max_stream_id_exclusive": min_stream_id,
-                "rows_inserted": rows_inserted + len(rows),
-            }
-
-            self._background_update_progress_txn(
-                txn, self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME, progress
-            )
-
-            return len(rows)
-
-        result = yield self.runInteraction(
-            self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME, reindex_txn
-        )
-
-        if not result:
-            yield self._end_background_update(self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME)
-
-        defer.returnValue(result)
-
-    @defer.inlineCallbacks
-    def _background_reindex_origin_server_ts(self, progress, batch_size):
-        target_min_stream_id = progress["target_min_stream_id_inclusive"]
-        max_stream_id = progress["max_stream_id_exclusive"]
-        rows_inserted = progress.get("rows_inserted", 0)
-
-        INSERT_CLUMP_SIZE = 1000
-
-        def reindex_search_txn(txn):
-            sql = (
-                "SELECT stream_ordering, event_id FROM events"
-                " WHERE ? <= stream_ordering AND stream_ordering < ?"
-                " ORDER BY stream_ordering DESC"
-                " LIMIT ?"
-            )
-
-            txn.execute(sql, (target_min_stream_id, max_stream_id, batch_size))
-
-            rows = txn.fetchall()
-            if not rows:
-                return 0
-
-            min_stream_id = rows[-1][0]
-            event_ids = [row[1] for row in rows]
-
-            rows_to_update = []
-
-            chunks = [event_ids[i : i + 100] for i in range(0, len(event_ids), 100)]
-            for chunk in chunks:
-                ev_rows = self._simple_select_many_txn(
-                    txn,
-                    table="event_json",
-                    column="event_id",
-                    iterable=chunk,
-                    retcols=["event_id", "json"],
-                    keyvalues={},
-                )
-
-                for row in ev_rows:
-                    event_id = row["event_id"]
-                    event_json = json.loads(row["json"])
-                    try:
-                        origin_server_ts = event_json["origin_server_ts"]
-                    except (KeyError, AttributeError):
-                        # If the event is missing a necessary field then
-                        # skip over it.
-                        continue
-
-                    rows_to_update.append((origin_server_ts, event_id))
-
-            sql = "UPDATE events SET origin_server_ts = ? WHERE event_id = ?"
-
-            for index in range(0, len(rows_to_update), INSERT_CLUMP_SIZE):
-                clump = rows_to_update[index : index + INSERT_CLUMP_SIZE]
-                txn.executemany(sql, clump)
-
-            progress = {
-                "target_min_stream_id_inclusive": target_min_stream_id,
-                "max_stream_id_exclusive": min_stream_id,
-                "rows_inserted": rows_inserted + len(rows_to_update),
-            }
-
-            self._background_update_progress_txn(
-                txn, self.EVENT_ORIGIN_SERVER_TS_NAME, progress
-            )
-
-            return len(rows_to_update)
-
-        result = yield self.runInteraction(
-            self.EVENT_ORIGIN_SERVER_TS_NAME, reindex_search_txn
-        )
-
-        if not result:
-            yield self._end_background_update(self.EVENT_ORIGIN_SERVER_TS_NAME)
-
-        defer.returnValue(result)
-
     def get_current_backfill_token(self):
         """The current minimum token that backfilled events have reached"""
         return -self._backfill_id_gen.get_current_token()
diff --git a/synapse/storage/events_bg_updates.py b/synapse/storage/events_bg_updates.py
new file mode 100644
index 0000000000..75c1935bf3
--- /dev/null
+++ b/synapse/storage/events_bg_updates.py
@@ -0,0 +1,401 @@
+# -*- coding: utf-8 -*-
+# Copyright 2019 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.
+
+import logging
+
+from six import text_type
+
+from canonicaljson import json
+
+from twisted.internet import defer
+
+from synapse.storage.background_updates import BackgroundUpdateStore
+
+logger = logging.getLogger(__name__)
+
+
+class EventsBackgroundUpdatesStore(BackgroundUpdateStore):
+
+    EVENT_ORIGIN_SERVER_TS_NAME = "event_origin_server_ts"
+    EVENT_FIELDS_SENDER_URL_UPDATE_NAME = "event_fields_sender_url"
+    DELETE_SOFT_FAILED_EXTREMITIES = "delete_soft_failed_extremities"
+
+    def __init__(self, db_conn, hs):
+        super(EventsBackgroundUpdatesStore, self).__init__(db_conn, hs)
+
+        self.register_background_update_handler(
+            self.EVENT_ORIGIN_SERVER_TS_NAME, self._background_reindex_origin_server_ts
+        )
+        self.register_background_update_handler(
+            self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME,
+            self._background_reindex_fields_sender,
+        )
+
+        self.register_background_index_update(
+            "event_contains_url_index",
+            index_name="event_contains_url_index",
+            table="events",
+            columns=["room_id", "topological_ordering", "stream_ordering"],
+            where_clause="contains_url = true AND outlier = false",
+        )
+
+        # an event_id index on event_search is useful for the purge_history
+        # api. Plus it means we get to enforce some integrity with a UNIQUE
+        # clause
+        self.register_background_index_update(
+            "event_search_event_id_idx",
+            index_name="event_search_event_id_idx",
+            table="event_search",
+            columns=["event_id"],
+            unique=True,
+            psql_only=True,
+        )
+
+        self.register_background_update_handler(
+            self.DELETE_SOFT_FAILED_EXTREMITIES,
+            self._cleanup_extremities_bg_update,
+        )
+
+    @defer.inlineCallbacks
+    def _background_reindex_fields_sender(self, progress, batch_size):
+        target_min_stream_id = progress["target_min_stream_id_inclusive"]
+        max_stream_id = progress["max_stream_id_exclusive"]
+        rows_inserted = progress.get("rows_inserted", 0)
+
+        INSERT_CLUMP_SIZE = 1000
+
+        def reindex_txn(txn):
+            sql = (
+                "SELECT stream_ordering, event_id, json FROM events"
+                " INNER JOIN event_json USING (event_id)"
+                " WHERE ? <= stream_ordering AND stream_ordering < ?"
+                " ORDER BY stream_ordering DESC"
+                " LIMIT ?"
+            )
+
+            txn.execute(sql, (target_min_stream_id, max_stream_id, batch_size))
+
+            rows = txn.fetchall()
+            if not rows:
+                return 0
+
+            min_stream_id = rows[-1][0]
+
+            update_rows = []
+            for row in rows:
+                try:
+                    event_id = row[1]
+                    event_json = json.loads(row[2])
+                    sender = event_json["sender"]
+                    content = event_json["content"]
+
+                    contains_url = "url" in content
+                    if contains_url:
+                        contains_url &= isinstance(content["url"], text_type)
+                except (KeyError, AttributeError):
+                    # If the event is missing a necessary field then
+                    # skip over it.
+                    continue
+
+                update_rows.append((sender, contains_url, event_id))
+
+            sql = "UPDATE events SET sender = ?, contains_url = ? WHERE event_id = ?"
+
+            for index in range(0, len(update_rows), INSERT_CLUMP_SIZE):
+                clump = update_rows[index : index + INSERT_CLUMP_SIZE]
+                txn.executemany(sql, clump)
+
+            progress = {
+                "target_min_stream_id_inclusive": target_min_stream_id,
+                "max_stream_id_exclusive": min_stream_id,
+                "rows_inserted": rows_inserted + len(rows),
+            }
+
+            self._background_update_progress_txn(
+                txn, self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME, progress
+            )
+
+            return len(rows)
+
+        result = yield self.runInteraction(
+            self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME, reindex_txn
+        )
+
+        if not result:
+            yield self._end_background_update(self.EVENT_FIELDS_SENDER_URL_UPDATE_NAME)
+
+        defer.returnValue(result)
+
+    @defer.inlineCallbacks
+    def _background_reindex_origin_server_ts(self, progress, batch_size):
+        target_min_stream_id = progress["target_min_stream_id_inclusive"]
+        max_stream_id = progress["max_stream_id_exclusive"]
+        rows_inserted = progress.get("rows_inserted", 0)
+
+        INSERT_CLUMP_SIZE = 1000
+
+        def reindex_search_txn(txn):
+            sql = (
+                "SELECT stream_ordering, event_id FROM events"
+                " WHERE ? <= stream_ordering AND stream_ordering < ?"
+                " ORDER BY stream_ordering DESC"
+                " LIMIT ?"
+            )
+
+            txn.execute(sql, (target_min_stream_id, max_stream_id, batch_size))
+
+            rows = txn.fetchall()
+            if not rows:
+                return 0
+
+            min_stream_id = rows[-1][0]
+            event_ids = [row[1] for row in rows]
+
+            rows_to_update = []
+
+            chunks = [event_ids[i : i + 100] for i in range(0, len(event_ids), 100)]
+            for chunk in chunks:
+                ev_rows = self._simple_select_many_txn(
+                    txn,
+                    table="event_json",
+                    column="event_id",
+                    iterable=chunk,
+                    retcols=["event_id", "json"],
+                    keyvalues={},
+                )
+
+                for row in ev_rows:
+                    event_id = row["event_id"]
+                    event_json = json.loads(row["json"])
+                    try:
+                        origin_server_ts = event_json["origin_server_ts"]
+                    except (KeyError, AttributeError):
+                        # If the event is missing a necessary field then
+                        # skip over it.
+                        continue
+
+                    rows_to_update.append((origin_server_ts, event_id))
+
+            sql = "UPDATE events SET origin_server_ts = ? WHERE event_id = ?"
+
+            for index in range(0, len(rows_to_update), INSERT_CLUMP_SIZE):
+                clump = rows_to_update[index : index + INSERT_CLUMP_SIZE]
+                txn.executemany(sql, clump)
+
+            progress = {
+                "target_min_stream_id_inclusive": target_min_stream_id,
+                "max_stream_id_exclusive": min_stream_id,
+                "rows_inserted": rows_inserted + len(rows_to_update),
+            }
+
+            self._background_update_progress_txn(
+                txn, self.EVENT_ORIGIN_SERVER_TS_NAME, progress
+            )
+
+            return len(rows_to_update)
+
+        result = yield self.runInteraction(
+            self.EVENT_ORIGIN_SERVER_TS_NAME, reindex_search_txn
+        )
+
+        if not result:
+            yield self._end_background_update(self.EVENT_ORIGIN_SERVER_TS_NAME)
+
+        defer.returnValue(result)
+
+    @defer.inlineCallbacks
+    def _cleanup_extremities_bg_update(self, progress, batch_size):
+        """Background update to clean out extremities that should have been
+        deleted previously.
+
+        Mainly used to deal with the aftermath of #5269.
+        """
+
+        # This works by first copying all existing forward extremities into the
+        # `_extremities_to_check` table at start up, and then checking each
+        # event in that table whether we have any descendants that are not
+        # soft-failed/rejected. If that is the case then we delete that event
+        # from the forward extremities table.
+        #
+        # For efficiency, we do this in batches by recursively pulling out all
+        # descendants of a batch until we find the non soft-failed/rejected
+        # events, i.e. the set of descendants whose chain of prev events back
+        # to the batch of extremities are all soft-failed or rejected.
+        # Typically, we won't find any such events as extremities will rarely
+        # have any descendants, but if they do then we should delete those
+        # extremities.
+
+        def _cleanup_extremities_bg_update_txn(txn):
+            # The set of extremity event IDs that we're checking this round
+            original_set = set()
+
+            # A dict[str, set[str]] of event ID to their prev events.
+            graph = {}
+
+            # The set of descendants of the original set that are not rejected
+            # nor soft-failed. Ancestors of these events should be removed
+            # from the forward extremities table.
+            non_rejected_leaves = set()
+
+            # Set of event IDs that have been soft failed, and for which we
+            # should check if they have descendants which haven't been soft
+            # failed.
+            soft_failed_events_to_lookup = set()
+
+            # First, we get `batch_size` events from the table, pulling out
+            # their successor events, if any, and the successor events'
+            # rejection status.
+            txn.execute(
+                """SELECT prev_event_id, event_id, internal_metadata,
+                    rejections.event_id IS NOT NULL, events.outlier
+                FROM (
+                    SELECT event_id AS prev_event_id
+                    FROM _extremities_to_check
+                    LIMIT ?
+                ) AS f
+                LEFT JOIN event_edges USING (prev_event_id)
+                LEFT JOIN events USING (event_id)
+                LEFT JOIN event_json USING (event_id)
+                LEFT JOIN rejections USING (event_id)
+                """, (batch_size,)
+            )
+
+            for prev_event_id, event_id, metadata, rejected, outlier in txn:
+                original_set.add(prev_event_id)
+
+                if not event_id or outlier:
+                    # Common case where the forward extremity doesn't have any
+                    # descendants.
+                    continue
+
+                graph.setdefault(event_id, set()).add(prev_event_id)
+
+                soft_failed = False
+                if metadata:
+                    soft_failed = json.loads(metadata).get("soft_failed")
+
+                if soft_failed or rejected:
+                    soft_failed_events_to_lookup.add(event_id)
+                else:
+                    non_rejected_leaves.add(event_id)
+
+            # Now we recursively check all the soft-failed descendants we
+            # found above in the same way, until we have nothing left to
+            # check.
+            while soft_failed_events_to_lookup:
+                # We only want to do 100 at a time, so we split given list
+                # into two.
+                batch = list(soft_failed_events_to_lookup)
+                to_check, to_defer = batch[:100], batch[100:]
+                soft_failed_events_to_lookup = set(to_defer)
+
+                sql = """SELECT prev_event_id, event_id, internal_metadata,
+                    rejections.event_id IS NOT NULL
+                    FROM event_edges
+                    INNER JOIN events USING (event_id)
+                    INNER JOIN event_json USING (event_id)
+                    LEFT JOIN rejections USING (event_id)
+                    WHERE
+                        prev_event_id IN (%s)
+                        AND NOT events.outlier
+                """ % (
+                    ",".join("?" for _ in to_check),
+                )
+                txn.execute(sql, to_check)
+
+                for prev_event_id, event_id, metadata, rejected in txn:
+                    if event_id in graph:
+                        # Already handled this event previously, but we still
+                        # want to record the edge.
+                        graph[event_id].add(prev_event_id)
+                        continue
+
+                    graph[event_id] = {prev_event_id}
+
+                    soft_failed = json.loads(metadata).get("soft_failed")
+                    if soft_failed or rejected:
+                        soft_failed_events_to_lookup.add(event_id)
+                    else:
+                        non_rejected_leaves.add(event_id)
+
+            # We have a set of non-soft-failed descendants, so we recurse up
+            # the graph to find all ancestors and add them to the set of event
+            # IDs that we can delete from forward extremities table.
+            to_delete = set()
+            while non_rejected_leaves:
+                event_id = non_rejected_leaves.pop()
+                prev_event_ids = graph.get(event_id, set())
+                non_rejected_leaves.update(prev_event_ids)
+                to_delete.update(prev_event_ids)
+
+            to_delete.intersection_update(original_set)
+
+            deleted = self._simple_delete_many_txn(
+                txn=txn,
+                table="event_forward_extremities",
+                column="event_id",
+                iterable=to_delete,
+                keyvalues={},
+            )
+
+            logger.info(
+                "Deleted %d forward extremities of %d checked, to clean up #5269",
+                deleted,
+                len(original_set),
+            )
+
+            if deleted:
+                # We now need to invalidate the caches of these rooms
+                rows = self._simple_select_many_txn(
+                    txn,
+                    table="events",
+                    column="event_id",
+                    iterable=to_delete,
+                    keyvalues={},
+                    retcols=("room_id",)
+                )
+                room_ids = set(row["room_id"] for row in rows)
+                for room_id in room_ids:
+                    txn.call_after(
+                        self.get_latest_event_ids_in_room.invalidate,
+                        (room_id,)
+                    )
+
+            self._simple_delete_many_txn(
+                txn=txn,
+                table="_extremities_to_check",
+                column="event_id",
+                iterable=original_set,
+                keyvalues={},
+            )
+
+            return len(original_set)
+
+        num_handled = yield self.runInteraction(
+            "_cleanup_extremities_bg_update", _cleanup_extremities_bg_update_txn,
+        )
+
+        if not num_handled:
+            yield self._end_background_update(self.DELETE_SOFT_FAILED_EXTREMITIES)
+
+            def _drop_table_txn(txn):
+                txn.execute("DROP TABLE _extremities_to_check")
+
+            yield self.runInteraction(
+                "_cleanup_extremities_bg_update_drop_table",
+                _drop_table_txn,
+            )
+
+        defer.returnValue(num_handled)
diff --git a/synapse/storage/schema/delta/54/delete_forward_extremities.sql b/synapse/storage/schema/delta/54/delete_forward_extremities.sql
new file mode 100644
index 0000000000..aa40f13da7
--- /dev/null
+++ b/synapse/storage/schema/delta/54/delete_forward_extremities.sql
@@ -0,0 +1,22 @@
+/* Copyright 2019 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.
+ */
+
+-- Start a background job to cleanup extremities that were incorrectly added
+-- by bug #5269.
+INSERT INTO background_updates (update_name, progress_json) VALUES
+  ('delete_soft_failed_extremities', '{}');
+
+DROP TABLE IF EXISTS _extremities_to_check;  -- To make this delta schema file idempotent.
+CREATE TABLE _extremities_to_check AS SELECT event_id FROM event_forward_extremities;
diff --git a/tests/storage/test_cleanup_extrems.py b/tests/storage/test_cleanup_extrems.py
new file mode 100644
index 0000000000..6dda66ecd3
--- /dev/null
+++ b/tests/storage/test_cleanup_extrems.py
@@ -0,0 +1,248 @@
+# -*- coding: utf-8 -*-
+# Copyright 2019 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.
+
+import os.path
+
+from synapse.api.constants import EventTypes
+from synapse.storage import prepare_database
+from synapse.types import Requester, UserID
+
+from tests.unittest import HomeserverTestCase
+
+
+class CleanupExtremBackgroundUpdateStoreTestCase(HomeserverTestCase):
+    """Test the background update to clean forward extremities table.
+    """
+
+    def prepare(self, reactor, clock, homeserver):
+        self.store = homeserver.get_datastore()
+        self.event_creator = homeserver.get_event_creation_handler()
+        self.room_creator = homeserver.get_room_creation_handler()
+
+        # Create a test user and room
+        self.user = UserID("alice", "test")
+        self.requester = Requester(self.user, None, False, None, None)
+        info = self.get_success(self.room_creator.create_room(self.requester, {}))
+        self.room_id = info["room_id"]
+
+    def create_and_send_event(self, soft_failed=False, prev_event_ids=None):
+        """Create and send an event.
+
+        Args:
+            soft_failed (bool): Whether to create a soft failed event or not
+            prev_event_ids (list[str]|None): Explicitly set the prev events,
+                or if None just use the default
+
+        Returns:
+            str: The new event's ID.
+        """
+        prev_events_and_hashes = None
+        if prev_event_ids:
+            prev_events_and_hashes = [[p, {}, 0] for p in prev_event_ids]
+
+        event, context = self.get_success(
+            self.event_creator.create_event(
+                self.requester,
+                {
+                    "type": EventTypes.Message,
+                    "room_id": self.room_id,
+                    "sender": self.user.to_string(),
+                    "content": {"body": "", "msgtype": "m.text"},
+                },
+                prev_events_and_hashes=prev_events_and_hashes,
+            )
+        )
+
+        if soft_failed:
+            event.internal_metadata.soft_failed = True
+
+        self.get_success(
+            self.event_creator.send_nonmember_event(self.requester, event, context)
+        )
+
+        return event.event_id
+
+    def add_extremity(self, event_id):
+        """Add the given event as an extremity to the room.
+        """
+        self.get_success(
+            self.store._simple_insert(
+                table="event_forward_extremities",
+                values={"room_id": self.room_id, "event_id": event_id},
+                desc="test_add_extremity",
+            )
+        )
+
+        self.store.get_latest_event_ids_in_room.invalidate((self.room_id,))
+
+    def run_background_update(self):
+        """Re run the background update to clean up the extremities.
+        """
+        # Make sure we don't clash with in progress updates.
+        self.assertTrue(self.store._all_done, "Background updates are still ongoing")
+
+        schema_path = os.path.join(
+            prepare_database.dir_path,
+            "schema",
+            "delta",
+            "54",
+            "delete_forward_extremities.sql",
+        )
+
+        def run_delta_file(txn):
+            prepare_database.executescript(txn, schema_path)
+
+        self.get_success(
+            self.store.runInteraction("test_delete_forward_extremities", run_delta_file)
+        )
+
+        # Ugh, have to reset this flag
+        self.store._all_done = False
+
+        while not self.get_success(self.store.has_completed_background_updates()):
+            self.get_success(self.store.do_next_background_update(100), by=0.1)
+
+    def test_soft_failed_extremities_handled_correctly(self):
+        """Test that extremities are correctly calculated in the presence of
+        soft failed events.
+
+        Tests a graph like:
+
+            A <- SF1 <- SF2 <- B
+
+        Where SF* are soft failed.
+        """
+
+        # Create the room graph
+        event_id_1 = self.create_and_send_event()
+        event_id_2 = self.create_and_send_event(True, [event_id_1])
+        event_id_3 = self.create_and_send_event(True, [event_id_2])
+        event_id_4 = self.create_and_send_event(False, [event_id_3])
+
+        # Check the latest events are as expected
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+
+        self.assertEqual(latest_event_ids, [event_id_4])
+
+    def test_basic_cleanup(self):
+        """Test that extremities are correctly calculated in the presence of
+        soft failed events.
+
+        Tests a graph like:
+
+            A <- SF1 <- B
+
+        Where SF* are soft failed, and with extremities of A and B
+        """
+        # Create the room graph
+        event_id_a = self.create_and_send_event()
+        event_id_sf1 = self.create_and_send_event(True, [event_id_a])
+        event_id_b = self.create_and_send_event(False, [event_id_sf1])
+
+        # Add the new extremity and check the latest events are as expected
+        self.add_extremity(event_id_a)
+
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+        self.assertEqual(set(latest_event_ids), set((event_id_a, event_id_b)))
+
+        # Run the background update and check it did the right thing
+        self.run_background_update()
+
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+        self.assertEqual(latest_event_ids, [event_id_b])
+
+    def test_chain_of_fail_cleanup(self):
+        """Test that extremities are correctly calculated in the presence of
+        soft failed events.
+
+        Tests a graph like:
+
+            A <- SF1 <- SF2 <- B
+
+        Where SF* are soft failed, and with extremities of A and B
+        """
+        # Create the room graph
+        event_id_a = self.create_and_send_event()
+        event_id_sf1 = self.create_and_send_event(True, [event_id_a])
+        event_id_sf2 = self.create_and_send_event(True, [event_id_sf1])
+        event_id_b = self.create_and_send_event(False, [event_id_sf2])
+
+        # Add the new extremity and check the latest events are as expected
+        self.add_extremity(event_id_a)
+
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+        self.assertEqual(set(latest_event_ids), set((event_id_a, event_id_b)))
+
+        # Run the background update and check it did the right thing
+        self.run_background_update()
+
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+        self.assertEqual(latest_event_ids, [event_id_b])
+
+    def test_forked_graph_cleanup(self):
+        r"""Test that extremities are correctly calculated in the presence of
+        soft failed events.
+
+        Tests a graph like, where time flows down the page:
+
+                A     B
+               / \   /
+              /   \ /
+            SF1   SF2
+             |     |
+            SF3    |
+           /   \   |
+           |    \  |
+           C     SF4
+
+        Where SF* are soft failed, and with them A, B and C marked as
+        extremities. This should resolve to B and C being marked as extremity.
+        """
+        # Create the room graph
+        event_id_a = self.create_and_send_event()
+        event_id_b = self.create_and_send_event()
+        event_id_sf1 = self.create_and_send_event(True, [event_id_a])
+        event_id_sf2 = self.create_and_send_event(True, [event_id_a, event_id_b])
+        event_id_sf3 = self.create_and_send_event(True, [event_id_sf1])
+        self.create_and_send_event(True, [event_id_sf2, event_id_sf3])  # SF4
+        event_id_c = self.create_and_send_event(False, [event_id_sf3])
+
+        # Add the new extremity and check the latest events are as expected
+        self.add_extremity(event_id_a)
+
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+        self.assertEqual(
+            set(latest_event_ids), set((event_id_a, event_id_b, event_id_c))
+        )
+
+        # Run the background update and check it did the right thing
+        self.run_background_update()
+
+        latest_event_ids = self.get_success(
+            self.store.get_latest_event_ids_in_room(self.room_id)
+        )
+        self.assertEqual(set(latest_event_ids), set([event_id_b, event_id_c]))