Peter Geoghegan <pg@heroku.com> writes:
> On Wed, Aug 20, 2014 at 11:08 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Is there a reason for these to behave inconsistently, and if not, which
>> behavior should we standardize on? Considering that you get NULL not an
>> error for extracting a nonexistent element from an object, I think there
>> is some case to be made for saying that returning NULL is the more
>> convenient behavior. Of course one can also argue for wanting this
>> operator to throw errors if the JSON structure doesn't match the
>> operation, but it seems like we've chosen to prefer being lax.
> I discussed this very issue with Andrew during development (I think
> that this happened to occur in private). My view was that since users
> will frequently use -> within expression indexes, it's best to have it
> return NULL for non-objects, rather than make them worry about the
> case where it'll be rejected, which is rather contrary to the spirit
> of jsonb (at least as a default behavior).
Hadn't thought of that angle, but it's a really good point. I'd certainly
rather be able to put an index on (jsoncol -> 'foo') without worrying
about what will happen if the column includes things that aren't objects.
> Andrew argued it was
> preferable to stick to the historic behavior of json operators. IMV,
> we should have both operators return NULL. They should be consistent,
> which implies changing the behavior of the existing json variants too,
> but I don't think that's a big problem.
None of these operators existed before 9.3, so I don't put a lot of stock
in the idea that their corner-case behaviors should be considered
sacrosanct already. But that will become the case pretty soon; if we
don't get it right in 9.4 it will arguably be too late.
regards, tom lane