Curt Sampson <cjs@cynic.net> writes:
> Well, should it then worry about read-ahead? On most OSes, it
> doesn't actually degenerate to 1-block random reads; it degerates
> to something along the lines of 8-block random reads.
Exactly. There is still an OS-provided benefit for sequential reading,
even if it's not as large as it might be in the absence of any other
activity. What this line of argument suggests is that random_page_cost
should be reduced in a heavily-loaded system ... but probably not to 1.0.
regards, tom lane