Re: patch: Add JSON datatype to PostgreSQL (GSoC, WIP)

It doesn't do particularly well on my previous example of [1,2,3].  It
comes out slightly shorter on ["a","b","c"] and better if the strings
need any escaping.  I don't think the float4 and float8 formats are
very useful; how could you be sure that the output was going to look
the same as the input?  Or alternatively that dumping a particular
object to text and reloading it will produce the same internal
representation?  I think it would be simpler to represent integers
using a string of digits; that way you can be assured of going from
text -> binary -> text without change.

Perhaps it would be enough to define the high two bits as follows: 00
= array, 01 = object, 10 = string, 11 = number/true/false/null.  The
next 2 bits specify how the length is stored.  00 = remaining 4 bits
store a length of up to 15 bytes, 01 = remaining 4 bits + 1 additional
byte store a 12-bit length of up to 4K, 10 = remaining 4 bits + 2
additional bytes store a 20-bit length of up to 1MB, 11 = 4 additional
bytes store a full 32-bit length word.  Then, the array, object, and
string representations can work as you've specified them.  Anything
else can be represented by itself, or perhaps we should say that
numbers represent themselves and true/false/null are represented by a
1-byte sequence, t/f/n (or perhaps we could define 111100{00,01,10} to
mean those values, since there's no obvious reason for the low bits to
be non-zero if a 4-bit length word ostensibly follows).

So [1,2,3] = 06 C1 '1' C1 '2' C1 '3' and ["a","b","c"] = 06 81 'a' 81 'b' 81 'c'

(I am still worried about the serialization/deserialization overhead
but that's a different issue.)