"David E. Wheeler" <da...@justatheory.com> writes:
> Bear in mind that `-` currently does both. Of the three current variants, the
> first two delete from an array by value:
> * jsonb - text: Deletes a key (and its value) from a JSON object, or matching
> string value(s) from a JSON array.
> * jsonb - text[] β jsonb: Deletes all matching keys or array elements from
> the left operand.
> * jsonb - integer β jsonb: Deletes the array element with specified index
> (negative integers count from the end). Throws an error if JSON value is not
> an array.
> Before I went and looked it up, I was also thinking this could use a
> different operator. But itβs already a bit overloaded, alas. So I could see
> the new behavior being:
> * jsonb - jsonb β jsonb: Deletes the array element with specified value.
> Throws an error if JSON value is not an array.
I fear that that would cause some problems. Consider
regression=# select '["foo", "bar"]'::jsonb - 'bar';
?column?
----------
["foo"]
(1 row)
Right now we resolve the unlabeled literal as type text.
But if jsonb - jsonb existed, we'd decide it's jsonb, thanks
to the heuristic that prefers same-type-as-the-other-input
(rule 2a at [1]). So it's pretty nearly certain that
adding jsonb - jsonb would break some existing queries;
or worse, silently cause them to do something different.
Maybe that's acceptable, but it's a demerit of this proposal.
regards, tom lane
[1] https://www.postgresql.org/docs/current/typeconv-oper.html
