diff options
Diffstat (limited to 'synapse/storage/events.py')
| -rw-r--r-- | synapse/storage/events.py | 157 |
1 files changed, 110 insertions, 47 deletions
diff --git a/synapse/storage/events.py b/synapse/storage/events.py index 8881b009df..919e855f3b 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" @@ -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 |