Batch process handling state groups

1 files changed, 21 insertions, 60 deletions

diff --git a/synapse/storage/events.py b/synapse/storage/events.py
index 0fb190530a..e4d0f8b1a9 100644
--- a/synapse/storage/events.py
+++ b/synapse/storage/events.py
@@ -37,6 +37,7 @@ from synapse.metrics.background_process_metrics import run_as_background_process
 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.async_helpers import ObservableDeferred
 from synapse.util.caches.descriptors import cached, cachedInlineCallbacks
@@ -203,7 +204,8 @@ 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"
 
@@ -1995,70 +1997,29 @@ 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 = []
-        unreferenced_state_groups = 0
-        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={},
-            )
-
-            for row in rows:
-                sg = row["state_group"]
-
-                if sg in state_groups_to_delete:
-                    # exclude state groups we are about to delete: no point in
-                    # updating them
-                    continue
+        referenced_state_groups =  set(sg for sg, in txn)
+        logger.info(
+            "[purge] found %i referenced state groups",
+            len(referenced_state_groups),
+        )
 
-                if not self._is_state_group_referenced(txn, sg):
-                    # Let's also delete unreferenced state groups while we're
-                    # here, since otherwise we'd need to de-delta them
-                    state_groups_to_delete.add(sg)
-                    unreferenced_state_groups += 1
-                    continue
+        logger.info("[purge] finding state groups that can be deleted")
 
-                remaining_state_groups.append(sg)
+        state_groups_to_delete, remaining_state_groups = self._find_unreferenced_groups(
+            txn, referenced_state_groups,
+        )
 
         logger.info(
-            "[purge] found %i extra unreferenced state groups to delete",
-            unreferenced_state_groups,
+            "[purge] found %i state groups to delete",
+            len(state_groups_to_delete),
         )
 
         logger.info(
@@ -2109,11 +2070,11 @@ 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")