"Matt Clark" <matt(at)ymogen(dot)net> writes:
> OK I'm definitely missing something here.
The point is that a big seqscan (either VACUUM or a plain table scan)
hits a lot of pages, and thereby tends to fill your cache with pages
that aren't actually likely to get hit again soon, perhaps pushing out
pages that will be needed again soon. This happens at both the
shared-buffer and kernel-disk-cache levels of caching.
It would be good to find some way to prevent big seqscans from
populating cache, but I don't know of any portable way to tell the OS
that we don't want it to cache a page we are reading.
regards, tom lane