> I think the logic you are referring to is the clock sweep buffer
> accounting
> scheme. That just makes sure that the most popular pages stay in the
> buffers. If your entire db fits in the buffer pool then it'll all get in
> there real fast.
Actually, I meant that in the case of a seq scan, PG will try to use just
a few buffers (a ring) in shared_buffers instead of thrashing the whole
buffers. But if there was actually a lot of free space in shared_buffers,
do the pages stay, or do they not ?