Simon Riggs <simon(at)2ndQuadrant(dot)com> writes:
> Surely the way to solve this is by having a new plan node that does a
> physical SeqScan of the index relation. It means we wouldn't preserve
> the sort order of the rows from the index, but that is just a plan
> cost issue.
> This is exactly what we do for VACUUM and it works faster there.
The reason that's okay for vacuum is that vacuum doesn't care if it
visits the same index tuple multiple times. It will not work for real
queries, unless you would like to lock out all concurrent inserts.
regards, tom lane