Speed up deleting of old rows in `event_push_actions` (#15531)
Enforce that we use index scans (rather than seq scans), which we also do for state queries. The reason to enforce this is that we can't correctly get PostgreSQL to understand the distribution of `stream_ordering` depends on `highlight`, and so it always defaults (on matrix.org) to sequential scans.pull/15533/head
							parent
							
								
									fc3a878220
								
							
						
					
					
						commit
						28ac1a1a91
					
				|  | @ -0,0 +1 @@ | |||
| Speed up deleting of old rows in `event_push_actions`. | ||||
|  | @ -1612,6 +1612,15 @@ class EventPushActionsWorkerStore(ReceiptsWorkerStore, StreamWorkerStore, SQLBas | |||
|             # deletes. | ||||
|             batch_size = self._rotate_count | ||||
| 
 | ||||
|             if isinstance(self.database_engine, PostgresEngine): | ||||
|                 # Temporarily disable sequential scans in this transaction. We | ||||
|                 # need to do this as the postgres statistics don't take into | ||||
|                 # account the `highlight = 0` part when estimating the | ||||
|                 # distribution of `stream_ordering`. I.e. since we keep old | ||||
|                 # highlight rows the query planner thinks there are way more old | ||||
|                 # rows to delete than there actually are. | ||||
|                 txn.execute("SET LOCAL enable_seqscan=off") | ||||
| 
 | ||||
|             txn.execute( | ||||
|                 """ | ||||
|                 SELECT stream_ordering FROM event_push_actions | ||||
|  |  | |||
		Loading…
	
		Reference in New Issue
	
	 Erik Johnston
						Erik Johnston