>>>>> "Tom" == Tom Lane <t...@sss.pgh.pa.us> writes:

 >> distance_orderable now returns true/false depending on the opclass,
 >> not just on the amcanorderbyop field. In order to do this, I've
 >> added an optional amproperty function to the AM api, which if it
 >> exists, gets first dibs on all property calls so it can override the
 >> result as it sees fit.

 Tom> Hmm, seems like for that case, it'd be easier to look into pg_amop
 Tom> and see if the opclass has any suitably-marked operators.

I thought about that, but it seemed like it could get painful. The
planner is working forwards from a known operator and matching it
against the index column, whereas we'd have to work backwards from the
opfamily, and there's no good existing index for this; in the presence
of binary-compatible types, I don't think even amoplefttype can be
assumed (e.g. a varchar column can be ordered by pg_trgm's <-> operator
which is declared for text). So it'd have to be the equivalent of:

  get index column's opclass oid
  look it up in pg_opclass to get opfamily
  for r in select * from pg_amop where amopfamily=? and amoppurpose='o'
    if r.amoplefttype is binary-coercible from the index column's type
      then return true

As opposed to what I have now, which is:

  get index column's opclass oid
  look it up in pg_opclass to get opfamily/opcintype
  result = SearchSysCacheExists4(AMPROCNUM, ...)  

(in theory this could produce a false positive if there's a distance
function but no actual operators to reference it, but I think that's the
opclass author's issue)

Andrew (irc:RhodiumToad)

Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:

Reply via email to