I ran a shorter version of David's script with just 6-9 attributes to
try to reproduce the problem area (spreadsheet with graph attached).
My test is also different in that I compare HEAD with just reverting
40af10b57. This shows a 60% increase in HEAD in runtime for 64MB
workmem and 64 byte tuples. It also shows a 20% regression for 32MB
workmem and 64 byte tuples.
I don't have anything to add to the discussion about whether something
needs to be done here for PG15. If anything, changing work_mem is an
easy to understand (although sometimes not practical) workaround.
--
John Naylor
EDB: http://www.enterprisedb.com