"Etsuro Fujita" <fujita.etsuro@lab.ntt.co.jp> writes:
> Agreed. However, I am concerned about the next comment in the current code:
> /*
> * Our generic assumption is that the index pages will be read
> * sequentially, so they cost seq_page_cost each, not random_page_cost.
> * ...
> I think this assumption is completely wrong, which has given me a motivation to
> propose a patch, though I am missing something.
Mph. It's pretty hard to argue that it's wrong without considering a
specific index implementation, which in practice would have a ton of
other details that need to be accounted for here. I don't have a strong
objection to changing the sample code to use random_page_cost instead,
but I doubt it will help anybody one way or another.
FWIW, the docs' sample code was an accurate transcription of what
genericcostestimate did at the time (8.1 era). I think the case we were
thinking of when that code was written was a recently-rebuilt btree
index, in which logically adjacent leaf pages would indeed often be
sequential.
regards, tom lane