Ba Jinsheng <bajinsheng@xxxxxxxxx> writes: >> It looks like the better plan involves a >> nestloop with inner indexscan on lineitem, which is something whose >> estimated cost depends enormously on random_page_cost. You've given >> us exactly zero detail about your test conditions, so it's hard to say >> more than that. > I used the default configuration in the file src/backend/utils/misc/postgresql.conf.sample > So the random_page_cost = 4.0 You're still admitting to nothing as to the hardware you are running this test on. However, 4.0 is a number we chose decades ago based on typical performance of spinning-rust storage. It's not very appropriate for SSD or similar storage -- numbers just a bit above 1 are probably the most appropriate thing for that kind of storage. (There are ongoing discussions about changing the setting's default value, but so far not backed by any great deal of hard evidence.) regards, tom lane
