Greetings,
* Jeff Janes (jeff.janes@gmail.com) wrote:
> In order to read 1409985 / 12839 = 109 rows per buffer page, the table must
> be extraordinarily well clustered on this index. That degree of clustering
> is going to steal much of the thunder from the index-only scan. But in my
> hands, it does still prefer the partial index with index-only scan by a
> cost estimate ratio of 3 to 1 (despite it actually being slightly slower)
> so I don't know why you don't get it being used.
Turns out to be because what was provided wasn't actually what was being
used- there's a domain in there and that seems to gum up the works and
make it so we don't consider the partial index as being something we can
use (see the discussion at the end of the other sub-thread).
Thanks,
Stephen