> I believe that the equal-key performance problem largely
> comes from the bogus way we've been handling duplicates, in particular
It comes from the way we look for page for new tuple - _bt_insertonpg tries
to avoid duplicate page splitting, i.e. if there is no space for new tuple
on the leftmost page, it reads next right page... etc...
> the fact that findsplitloc has to worry about choosing a "legal split
> point" for a series of duplicates. Once we get rid of that, I think
> findsplitloc can use a binary search.
Vadim