Fix order of partial state tables when purging (#15068)

* Fix order of partial state tables when purging

`partial_state_rooms` has an FK on `events` pointing to the join event we
get from `/send_join`, so we must delete from that table before deleting
from `events`.

**NB:** It would be nice to cancel any resync processes for the room
being purged. We do not do this at present. To do so reliably we'd need
an internal HTTP "replication" endpoint, because the worker doing the
resync process may be different to that handling the purge request.

The first time the resync process tries to write data after the deletion
it will fail because we have deleted necessary data e.g. auth
events. AFAICS it will not retry the resync, so the only downside to
not cancelling the resync is a scary-looking traceback.

(This is presumably extremely race-sensitive.)

* Changelog

* admist(?) -> between

* Warn about a race

* Fix typo, thanks Sean

Co-authored-by: Sean Quah <8349537+squahtx@users.noreply.github.com>

---------

Co-authored-by: Sean Quah <8349537+squahtx@users.noreply.github.com>
This commit is contained in:
David Robertson 2023-02-14 23:42:29 +00:00 committed by GitHub
parent 42aea0d8af
commit 06ba71083e
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 10 additions and 2 deletions

1
changelog.d/15068.bugfix Normal file
View file

@ -0,0 +1 @@
Fix a bug introduced in Synapse 1.76.0 where partially-joined rooms could not be deleted using the [purge room API](https://matrix-org.github.io/synapse/latest/admin_api/rooms.html#delete-room-api).

View file

@ -1880,6 +1880,11 @@ class FederationHandler:
logger.info("Updating current state for %s", room_id)
# TODO(faster_joins): notify workers in notify_room_un_partial_stated
# https://github.com/matrix-org/synapse/issues/12994
#
# NB: there's a potential race here. If room is purged just before we
# call this, we _might_ end up inserting rows into current_state_events.
# (The logic is hard to chase through.) We think this is fine, but if
# not the HS admin should purge the room again.
await self.state_handler.update_current_state(room_id)
logger.info("Handling any pending device list updates")

View file

@ -420,12 +420,14 @@ class PurgeEventsStore(StateGroupWorkerStore, CacheInvalidationWorkerStore):
"event_push_actions",
"event_search",
"event_failed_pull_attempts",
# Note: the partial state tables have foreign keys between each other, and to
# `events` and `rooms`. We need to delete from them in the right order.
"partial_state_events",
"partial_state_rooms_servers",
"partial_state_rooms",
"events",
"federation_inbound_events_staging",
"local_current_membership",
"partial_state_rooms_servers",
"partial_state_rooms",
"receipts_graph",
"receipts_linearized",
"room_aliases",