Heikki Linnakangas <hlinnaka@iki.fi> writes:
> On 23/05/18 19:25, Tom Lane wrote:
>> To make this
>> patch safe, I think you'd need to grovel through the subquery and make
>> sure that the parameters are only used as inputs to operators that belong
>> to the type's default btree or hash opfamily. (Many other cases would
>> work in practice, but we have no semantic knowledge that would let us be
>> sure of that.)
> Hmm. First thing that comes to mind is to use the raw bytes as cache
> key, only treating Datums as equal if their binary representation is
> identical.
Ah. That would work, though it'd make the number of subquery executions
even less predictable (since some logically-equal values would compare
as physically unequal).
regards, tom lane