Merge remote-tracking branch 'origin/develop' into dbkr/e2e_backup_versions_are_numbers

8 files changed, 274 insertions, 65 deletions

diff --git a/synapse/storage/devices.py b/synapse/storage/devices.py
index 62497ab63f..ecdab34e7d 100644
--- a/synapse/storage/devices.py
+++ b/synapse/storage/devices.py
@@ -22,14 +22,19 @@ from twisted.internet import defer
 
 from synapse.api.errors import StoreError
 from synapse.metrics.background_process_metrics import run_as_background_process
+from synapse.storage.background_updates import BackgroundUpdateStore
 from synapse.util.caches.descriptors import cached, cachedInlineCallbacks, cachedList
 
-from ._base import Cache, SQLBaseStore, db_to_json
+from ._base import Cache, db_to_json
 
 logger = logging.getLogger(__name__)
 
+DROP_DEVICE_LIST_STREAMS_NON_UNIQUE_INDEXES = (
+    "drop_device_list_streams_non_unique_indexes"
+)
 
-class DeviceStore(SQLBaseStore):
+
+class DeviceStore(BackgroundUpdateStore):
     def __init__(self, db_conn, hs):
         super(DeviceStore, self).__init__(db_conn, hs)
 
@@ -52,6 +57,30 @@ class DeviceStore(SQLBaseStore):
             columns=["user_id", "device_id"],
         )
 
+        # create a unique index on device_lists_remote_cache
+        self.register_background_index_update(
+            "device_lists_remote_cache_unique_idx",
+            index_name="device_lists_remote_cache_unique_id",
+            table="device_lists_remote_cache",
+            columns=["user_id", "device_id"],
+            unique=True,
+        )
+
+        # And one on device_lists_remote_extremeties
+        self.register_background_index_update(
+            "device_lists_remote_extremeties_unique_idx",
+            index_name="device_lists_remote_extremeties_unique_idx",
+            table="device_lists_remote_extremeties",
+            columns=["user_id"],
+            unique=True,
+        )
+
+        # once they complete, we can remove the old non-unique indexes.
+        self.register_background_update_handler(
+            DROP_DEVICE_LIST_STREAMS_NON_UNIQUE_INDEXES,
+            self._drop_device_list_streams_non_unique_indexes,
+        )
+
     @defer.inlineCallbacks
     def store_device(self, user_id, device_id,
                      initial_device_display_name):
@@ -239,7 +268,19 @@ class DeviceStore(SQLBaseStore):
 
     def update_remote_device_list_cache_entry(self, user_id, device_id, content,
                                               stream_id):
-        """Updates a single user's device in the cache.
+        """Updates a single device in the cache of a remote user's devicelist.
+
+        Note: assumes that we are the only thread that can be updating this user's
+        device list.
+
+        Args:
+            user_id (str): User to update device list for
+            device_id (str): ID of decivice being updated
+            content (dict): new data on this device
+            stream_id (int): the version of the device list
+
+        Returns:
+            Deferred[None]
         """
         return self.runInteraction(
             "update_remote_device_list_cache_entry",
@@ -272,7 +313,11 @@ class DeviceStore(SQLBaseStore):
                 },
                 values={
                     "content": json.dumps(content),
-                }
+                },
+
+                # we don't need to lock, because we assume we are the only thread
+                # updating this user's devices.
+                lock=False,
             )
 
         txn.call_after(self._get_cached_user_device.invalidate, (user_id, device_id,))
@@ -289,11 +334,26 @@ class DeviceStore(SQLBaseStore):
             },
             values={
                 "stream_id": stream_id,
-            }
+            },
+
+            # again, we can assume we are the only thread updating this user's
+            # extremity.
+            lock=False,
         )
 
     def update_remote_device_list_cache(self, user_id, devices, stream_id):
-        """Replace the cache of the remote user's devices.
+        """Replace the entire cache of the remote user's devices.
+
+        Note: assumes that we are the only thread that can be updating this user's
+        device list.
+
+        Args:
+            user_id (str): User to update device list for
+            devices (list[dict]): list of device objects supplied over federation
+            stream_id (int): the version of the device list
+
+        Returns:
+            Deferred[None]
         """
         return self.runInteraction(
             "update_remote_device_list_cache",
@@ -338,7 +398,11 @@ class DeviceStore(SQLBaseStore):
             },
             values={
                 "stream_id": stream_id,
-            }
+            },
+
+            # we don't need to lock, because we can assume we are the only thread
+            # updating this user's extremity.
+            lock=False,
         )
 
     def get_devices_by_remote(self, destination, from_stream_id):
@@ -722,3 +786,19 @@ class DeviceStore(SQLBaseStore):
             "_prune_old_outbound_device_pokes",
             _prune_txn,
         )
+
+    @defer.inlineCallbacks
+    def _drop_device_list_streams_non_unique_indexes(self, progress, batch_size):
+        def f(conn):
+            txn = conn.cursor()
+            txn.execute(
+                "DROP INDEX IF EXISTS device_lists_remote_cache_id"
+            )
+            txn.execute(
+                "DROP INDEX IF EXISTS device_lists_remote_extremeties_id"
+            )
+            txn.close()
+
+        yield self.runWithConnection(f)
+        yield self._end_background_update(DROP_DEVICE_LIST_STREAMS_NON_UNIQUE_INDEXES)
+        defer.returnValue(1)
diff --git a/synapse/storage/end_to_end_keys.py b/synapse/storage/end_to_end_keys.py
index 1f1721e820..2a0f6cfca9 100644
--- a/synapse/storage/end_to_end_keys.py
+++ b/synapse/storage/end_to_end_keys.py
@@ -40,7 +40,10 @@ class EndToEndKeyStore(SQLBaseStore):
                 allow_none=True,
             )
 
-            new_key_json = encode_canonical_json(device_keys)
+            # In py3 we need old_key_json to match new_key_json type. The DB
+            # returns unicode while encode_canonical_json returns bytes.
+            new_key_json = encode_canonical_json(device_keys).decode("utf-8")
+
             if old_key_json == new_key_json:
                 return False
 
diff --git a/synapse/storage/event_federation.py b/synapse/storage/event_federation.py
index 3faca2a042..d3b9dea1d6 100644
--- a/synapse/storage/event_federation.py
+++ b/synapse/storage/event_federation.py
@@ -477,7 +477,7 @@ class EventFederationStore(EventFederationWorkerStore):
                     "is_state": False,
                 }
                 for ev in events
-                for e_id, _ in ev.prev_events
+                for e_id in ev.prev_event_ids()
             ],
         )
 
@@ -510,7 +510,7 @@ class EventFederationStore(EventFederationWorkerStore):
 
         txn.executemany(query, [
             (e_id, ev.room_id, e_id, ev.room_id, e_id, ev.room_id, False)
-            for ev in events for e_id, _ in ev.prev_events
+            for ev in events for e_id in ev.prev_event_ids()
             if not ev.internal_metadata.is_outlier()
         ])
 
diff --git a/synapse/storage/events.py b/synapse/storage/events.py
index 8881b009df..2047110b1d 100644
--- a/synapse/storage/events.py
+++ b/synapse/storage/events.py
@@ -38,6 +38,7 @@ from synapse.state import StateResolutionStore
 from synapse.storage.background_updates import BackgroundUpdateStore
 from synapse.storage.event_federation import EventFederationStore
 from synapse.storage.events_worker import EventsWorkerStore
+from synapse.storage.state import StateGroupWorkerStore
 from synapse.types import RoomStreamToken, get_domain_from_id
 from synapse.util import batch_iter
 from synapse.util.async_helpers import ObservableDeferred
@@ -205,7 +206,8 @@ def _retry_on_integrity_error(func):
 
 # inherits from EventFederationStore so that we can call _update_backward_extremities
 # and _handle_mult_prev_events (though arguably those could both be moved in here)
-class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore):
+class EventsStore(StateGroupWorkerStore, EventFederationStore, EventsWorkerStore,
+                  BackgroundUpdateStore):
     EVENT_ORIGIN_SERVER_TS_NAME = "event_origin_server_ts"
     EVENT_FIELDS_SENDER_URL_UPDATE_NAME = "event_fields_sender_url"
 
@@ -414,7 +416,7 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
                             )
                             if len_1:
                                 all_single_prev_not_state = all(
-                                    len(event.prev_events) == 1
+                                    len(event.prev_event_ids()) == 1
                                     and not event.is_state()
                                     for event, ctx in ev_ctx_rm
                                 )
@@ -438,7 +440,7 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
                                 # guess this by looking at the prev_events and checking
                                 # if they match the current forward extremities.
                                 for ev, _ in ev_ctx_rm:
-                                    prev_event_ids = set(e for e, _ in ev.prev_events)
+                                    prev_event_ids = set(ev.prev_event_ids())
                                     if latest_event_ids == prev_event_ids:
                                         state_delta_reuse_delta_counter.inc()
                                         break
@@ -549,7 +551,7 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
         result.difference_update(
             e_id
             for event in new_events
-            for e_id, _ in event.prev_events
+            for e_id in event.prev_event_ids()
         )
 
         # Finally, remove any events which are prev_events of any existing events.
@@ -867,7 +869,7 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
                     "auth_id": auth_id,
                 }
                 for event, _ in events_and_contexts
-                for auth_id, _ in event.auth_events
+                for auth_id in event.auth_event_ids()
                 if event.is_state()
             ],
         )
@@ -2034,55 +2036,37 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
 
         logger.info("[purge] finding redundant state groups")
 
-        # Get all state groups that are only referenced by events that are
-        # to be deleted.
-        # This works by first getting state groups that we may want to delete,
-        # joining against event_to_state_groups to get events that use that
-        # state group, then left joining against events_to_purge again. Any
-        # state group where the left join produce *no nulls* are referenced
-        # only by events that are going to be purged.
+        # Get all state groups that are referenced by events that are to be
+        # deleted. We then go and check if they are referenced by other events
+        # or state groups, and if not we delete them.
         txn.execute("""
-            SELECT state_group FROM
-            (
-                SELECT DISTINCT state_group FROM events_to_purge
-                INNER JOIN event_to_state_groups USING (event_id)
-            ) AS sp
-            INNER JOIN event_to_state_groups USING (state_group)
-            LEFT JOIN events_to_purge AS ep USING (event_id)
-            GROUP BY state_group
-            HAVING SUM(CASE WHEN ep.event_id IS NULL THEN 1 ELSE 0 END) = 0
+            SELECT DISTINCT state_group FROM events_to_purge
+            INNER JOIN event_to_state_groups USING (event_id)
         """)
 
-        state_rows = txn.fetchall()
-        logger.info("[purge] found %i redundant state groups", len(state_rows))
-
-        # make a set of the redundant state groups, so that we can look them up
-        # efficiently
-        state_groups_to_delete = set([sg for sg, in state_rows])
-
-        # Now we get all the state groups that rely on these state groups
-        logger.info("[purge] finding state groups which depend on redundant"
-                    " state groups")
-        remaining_state_groups = []
-        for i in range(0, len(state_rows), 100):
-            chunk = [sg for sg, in state_rows[i:i + 100]]
-            # look for state groups whose prev_state_group is one we are about
-            # to delete
-            rows = self._simple_select_many_txn(
-                txn,
-                table="state_group_edges",
-                column="prev_state_group",
-                iterable=chunk,
-                retcols=["state_group"],
-                keyvalues={},
-            )
-            remaining_state_groups.extend(
-                row["state_group"] for row in rows
+        referenced_state_groups = set(sg for sg, in txn)
+        logger.info(
+            "[purge] found %i referenced state groups",
+            len(referenced_state_groups),
+        )
 
-                # exclude state groups we are about to delete: no point in
-                # updating them
-                if row["state_group"] not in state_groups_to_delete
+        logger.info("[purge] finding state groups that can be deleted")
+
+        state_groups_to_delete, remaining_state_groups = (
+            self._find_unreferenced_groups_during_purge(
+                txn, referenced_state_groups,
             )
+        )
+
+        logger.info(
+            "[purge] found %i state groups to delete",
+            len(state_groups_to_delete),
+        )
+
+        logger.info(
+            "[purge] de-delta-ing %i remaining state groups",
+            len(remaining_state_groups),
+        )
 
         # Now we turn the state groups that reference to-be-deleted state
         # groups to non delta versions.
@@ -2127,11 +2111,11 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
         logger.info("[purge] removing redundant state groups")
         txn.executemany(
             "DELETE FROM state_groups_state WHERE state_group = ?",
-            state_rows
+            ((sg,) for sg in state_groups_to_delete),
         )
         txn.executemany(
             "DELETE FROM state_groups WHERE id = ?",
-            state_rows
+            ((sg,) for sg in state_groups_to_delete),
         )
 
         logger.info("[purge] removing events from event_to_state_groups")
@@ -2227,6 +2211,85 @@ class EventsStore(EventFederationStore, EventsWorkerStore, BackgroundUpdateStore
 
         logger.info("[purge] done")
 
+    def _find_unreferenced_groups_during_purge(self, txn, state_groups):
+        """Used when purging history to figure out which state groups can be
+        deleted and which need to be de-delta'ed (due to one of its prev groups
+        being scheduled for deletion).
+
+        Args:
+            txn
+            state_groups (set[int]): Set of state groups referenced by events
+                that are going to be deleted.
+
+        Returns:
+            tuple[set[int], set[int]]: The set of state groups that can be
+            deleted and the set of state groups that need to be de-delta'ed
+        """
+        # Graph of state group -> previous group
+        graph = {}
+
+        # Set of events that we have found to be referenced by events
+        referenced_groups = set()
+
+        # Set of state groups we've already seen
+        state_groups_seen = set(state_groups)
+
+        # Set of state groups to handle next.
+        next_to_search = set(state_groups)
+        while next_to_search:
+            # We bound size of groups we're looking up at once, to stop the
+            # SQL query getting too big
+            if len(next_to_search) < 100:
+                current_search = next_to_search
+                next_to_search = set()
+            else:
+                current_search = set(itertools.islice(next_to_search, 100))
+                next_to_search -= current_search
+
+            # Check if state groups are referenced
+            sql = """
+                SELECT DISTINCT state_group FROM event_to_state_groups
+                LEFT JOIN events_to_purge AS ep USING (event_id)
+                WHERE state_group IN (%s) AND ep.event_id IS NULL
+            """ % (",".join("?" for _ in current_search),)
+            txn.execute(sql, list(current_search))
+
+            referenced = set(sg for sg, in txn)
+            referenced_groups |= referenced
+
+            # We don't continue iterating up the state group graphs for state
+            # groups that are referenced.
+            current_search -= referenced
+
+            rows = self._simple_select_many_txn(
+                txn,
+                table="state_group_edges",
+                column="prev_state_group",
+                iterable=current_search,
+                keyvalues={},
+                retcols=("prev_state_group", "state_group",),
+            )
+
+            prevs = set(row["state_group"] for row in rows)
+            # We don't bother re-handling groups we've already seen
+            prevs -= state_groups_seen
+            next_to_search |= prevs
+            state_groups_seen |= prevs
+
+            for row in rows:
+                # Note: Each state group can have at most one prev group
+                graph[row["state_group"]] = row["prev_state_group"]
+
+        to_delete = state_groups_seen - referenced_groups
+
+        to_dedelta = set()
+        for sg in referenced_groups:
+            prev_sg = graph.get(sg)
+            if prev_sg and prev_sg in to_delete:
+                to_dedelta.add(sg)
+
+        return to_delete, to_dedelta
+
     @defer.inlineCallbacks
     def is_event_after(self, event_id1, event_id2):
         """Returns True if event_id1 is after event_id2 in the stream
diff --git a/synapse/storage/schema/delta/40/device_list_streams.sql b/synapse/storage/schema/delta/40/device_list_streams.sql
index 54841b3843..dd6dcb65f1 100644
--- a/synapse/storage/schema/delta/40/device_list_streams.sql
+++ b/synapse/storage/schema/delta/40/device_list_streams.sql
@@ -20,9 +20,6 @@ CREATE TABLE device_lists_remote_cache (
     content TEXT NOT NULL
 );
 
-CREATE INDEX device_lists_remote_cache_id ON device_lists_remote_cache(user_id, device_id);
-
-
 -- The last update we got for a user. Empty if we're not receiving updates for
 -- that user.
 CREATE TABLE device_lists_remote_extremeties (
@@ -30,7 +27,11 @@ CREATE TABLE device_lists_remote_extremeties (
     stream_id TEXT NOT NULL
 );
 
-CREATE INDEX device_lists_remote_extremeties_id ON device_lists_remote_extremeties(user_id, stream_id);
+-- we used to create non-unique indexes on these tables, but as of update 52 we create
+-- unique indexes concurrently:
+--
+-- CREATE INDEX device_lists_remote_cache_id ON device_lists_remote_cache(user_id, device_id);
+-- CREATE INDEX device_lists_remote_extremeties_id ON device_lists_remote_extremeties(user_id, stream_id);
 
 
 -- Stream of device lists updates. Includes both local and remotes
diff --git a/synapse/storage/schema/delta/52/add_event_to_state_group_index.sql b/synapse/storage/schema/delta/52/add_event_to_state_group_index.sql
new file mode 100644
index 0000000000..91e03d13e1
--- /dev/null
+++ b/synapse/storage/schema/delta/52/add_event_to_state_group_index.sql
@@ -0,0 +1,19 @@
+/* Copyright 2018 New Vector Ltd
+ *
+ * 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.
+ */
+
+-- This is needed to efficiently check for unreferenced state groups during
+-- purge. Added events_to_state_group(state_group) index
+INSERT into background_updates (update_name, progress_json)
+    VALUES ('event_to_state_groups_sg_index', '{}');
diff --git a/synapse/storage/schema/delta/52/device_list_streams_unique_idx.sql b/synapse/storage/schema/delta/52/device_list_streams_unique_idx.sql
new file mode 100644
index 0000000000..bfa49e6f92
--- /dev/null
+++ b/synapse/storage/schema/delta/52/device_list_streams_unique_idx.sql
@@ -0,0 +1,36 @@
+/* Copyright 2018 New Vector Ltd
+ *
+ * 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.
+ */
+
+-- register a background update which will create a unique index on
+-- device_lists_remote_cache
+INSERT into background_updates (update_name, progress_json)
+    VALUES ('device_lists_remote_cache_unique_idx', '{}');
+
+-- and one on device_lists_remote_extremeties
+INSERT into background_updates (update_name, progress_json, depends_on)
+    VALUES (
+        'device_lists_remote_extremeties_unique_idx', '{}',
+
+        -- doesn't really depend on this, but we need to make sure both happen
+        -- before we drop the old indexes.
+        'device_lists_remote_cache_unique_idx'
+    );
+
+-- once they complete, we can drop the old indexes.
+INSERT into background_updates (update_name, progress_json, depends_on)
+    VALUES (
+        'drop_device_list_streams_non_unique_indexes', '{}',
+        'device_lists_remote_extremeties_unique_idx'
+    );
diff --git a/synapse/storage/state.py b/synapse/storage/state.py
index ef65929bb2..d737bd6778 100644
--- a/synapse/storage/state.py
+++ b/synapse/storage/state.py
@@ -1257,6 +1257,7 @@ class StateStore(StateGroupWorkerStore, BackgroundUpdateStore):
     STATE_GROUP_DEDUPLICATION_UPDATE_NAME = "state_group_state_deduplication"
     STATE_GROUP_INDEX_UPDATE_NAME = "state_group_state_type_index"
     CURRENT_STATE_INDEX_UPDATE_NAME = "current_state_members_idx"
+    EVENT_STATE_GROUP_INDEX_UPDATE_NAME = "event_to_state_groups_sg_index"
 
     def __init__(self, db_conn, hs):
         super(StateStore, self).__init__(db_conn, hs)
@@ -1275,6 +1276,12 @@ class StateStore(StateGroupWorkerStore, BackgroundUpdateStore):
             columns=["state_key"],
             where_clause="type='m.room.member'",
         )
+        self.register_background_index_update(
+            self.EVENT_STATE_GROUP_INDEX_UPDATE_NAME,
+            index_name="event_to_state_groups_sg_index",
+            table="event_to_state_groups",
+            columns=["state_group"],
+        )
 
     def _store_event_state_mappings_txn(self, txn, events_and_contexts):
         state_groups = {}