On Fri, Feb 2, 2018 at 4:53 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:
>> BTW wouldn't it be possible to derive "traditional" proof in relational
>> algebra, similarly to other transforms?
>
> Perhaps. The patch depends on CTID, but you could probably model that
> as a primary key in a proof.
I'll remind you that commit b648b703 made the TID sort performed by
CREATE INDEX CONCURRENTLY over 3 times faster in cases where the sort
completes in memory. The simplest way to get a significant portion of
that benefit for your approach with sort+unique on CTID would be to
add sortsupport with abbreviated keys to the TID btree opclass.
--
Peter Geoghegan