Re: [PERFORM] Re: querying with index on jsonb slower than standard column. Why?

Andrew Dunstan <andrew@dunslane.net> writes:
> On 12/12/2014 04:44 PM, Tom Lane wrote:
>> Well, it'd be easy to fix if we were willing to invent distinct operators
>> depending on which type you wanted out (perhaps ->> for text output as
>> today, add ->># for numeric output, etc).

> That was my immediate reaction. Not sure about the operator name. I'd
> tentatively suggest -># (taking an int or text argument) and #># taking
> a text[] argument, both returning numeric, and erroring out if the value
> is a string, boolean, object or array.

>> The usability issue could be fixed by teaching the planner to fold a
>> construct like (jsonb ->> 'foo')::numeric into (jsonb ->># 'foo').
>> But I'm not sure how we do that except in a really ugly and ad-hoc
>> fashion.

> I would be inclined to add the operator and see how cumbersome people
> find it. I suspect in many cases it might be sufficient.

We can't just add the operator and worry about usability later;
if we're thinking we might want to introduce such an automatic
transformation, we have to be sure the new operator is defined in a
way that allows the transformation to not change any semantics.
What that means in this case is that if (jsonb ->> 'foo')::numeric
would have succeeded, (jsonb ->># 'foo') has to succeed; which means
it'd better be willing to attempt conversion of string values to
numeric, not just throw an error on sight.

            regards, tom lane