On 03/13/2013 08:17 AM, Robert Haas wrote:
> On Fri, Mar 8, 2013 at 4:42 PM, Andrew Dunstan <andrew@dunslane.net> wrote:
>>> So my order of preference for the options would be:
>>>
>>> 1. Have the JSON type collapse objects so the last instance of a key wins
>>> and is actually stored
>>>
>>> 2. Throw an error when a JSON type has duplicate keys
>>>
>>> 3. Have the accessors find the last instance of a key and return that
>>> value
>>>
>>> 4. Let things remain as they are now
>>>
>>> On second though, I don't like 4 at all. It means that the JSON type
>>> things a value is valid while the accessor does not. They contradict one
>>> another.
>> You can forget 1. We are not going to have the parser collapse anything.
>> Either the JSON it gets is valid or it's not. But the parser isn't going to
>> try to MAKE it valid.
> Why not? Because it's the wrong thing to do, or because it would be slower?
>
> What I think is tricky here is that there's more than one way to
> conceptualize what the JSON data type really is. Is it a key-value
> store of sorts, or just a way to store text values that meet certain
> minimalist syntactic criteria? I had imagined it as the latter, in
> which case normalization isn't sensible. But if you think of it the
> first way, then normalization is not only sensible, but almost
> obligatory. For example, we don't feel bad about this:
>
> rhaas=# select '1e1'::numeric;
> numeric
> ---------
> 10
> (1 row)
>
> I think Andrew and I had envisioned this as basically a text data type
> that enforces some syntax checking on its input, hence the current
> design. But I'm not sure that's the ONLY sensible design.
>
I think we've moved on from this point, because a) other implementations
allow duplicate keys, b) it's trivially easy to make Postgres generate
such json, and c) there is some dispute about exactly what the spec
mandates.
I'll be posting a revised patch shortly that doesn't error out but
simply uses the value for the later key lexically.
cheers
andrew