Re: Support for RANGE ... PRECEDING windows in OVER

Robert Haas <robertmhaas@gmail.com> writes:
> On Mon, Jul 1, 2013 at 3:28 PM, Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>> Currently, all operator classes are tied to access methods.  Since
>> nobody seems to have any great idea about creating an access method that
>> requires addition and subtraction, would it make sense to have operator
>> classes that exist solely to support keeping track of such operators for
>> the various datatypes?

We certainly could envision adding "+" and "-" items to btree opfamilies,
with the proviso that they'd have to be optional so as not to break
existing extensions that create btree opfamilies.  If we went that way,
I'd suggest that what we actually add to the definition is *not*
operators, but functions --- that is, specify that function slots 2 and
3 can contain addition (resp subtraction) functions that are compatible
with the family's ordering behavior.  Otherwise you'd have the planner
trying to match WHERE clauses to the extra operators, which is somewhere
between wrong and dangerous.  But a function that isn't actually called
by the index AM is not going to pose a hazard of being misapplied.
Besides, it'd likely be easier not harder for the window-function
machinery to work with a function than an operator.

> I suppose if we really wanted to do this, it would make more sense to
> have a new kind of object, maybe CREATE TYPE INTERFACE, rather than
> shoehorning it into the operator class machinery.  It seems like a
> fairly heavyweight solution, however.

Yeah, there's something to be said for not wedging this into the index
opclass infrastructure.  I'd be happier about building such new
infrastructure if we could unify this requirement with Peter's
"transforms" feature, which also seems to need more datatype-related
knowledge than we have in the catalogs now.
        regards, tom lane