Hi hackers,
Right now cost of index-only scan is using `random_page_cost`.
Certainly for point selects we really have random access pattern, but
queries like "select count(*) from hits" access pattern is more or less
sequential:
we are iterating through subsequent leaf B-Tree pages. As far as
default value of `random_page_cost` is 4 times larger than `seq_page_cost`
it may force Postgres optimizer to choose sequential scan, while
index-only scan is usually much faster in this case.
Can we do something here to provide more accurate cost estimation?