Is attribute order stable? Obviously, order is not preserved, but if the order changes on subsequent accesses, this causes problems if you ever serve content directly from a jsonb field without sorting the attributes manually.
[0] http://json.org/
EDIT: Well, now four of us have replied at the same time saying the same thing, I'd say this topic is now well covered.
Anyway, it should be possible to write a simple serializer on your own - a trivial PL/V8 function should suffice, I guess. And then you can define a CAST using that function (but generally adding casts is a bit dangerous, as it may have unexpected consequences).
"An object is an unordered set of name/value pairs."
1. Cache control using etags. If the content changes by a single byte, even if semantically identical, the etag should change. Hence '{"foo":1,"bar":2}' is not equivalent to '{"bar":2,"foo":1}'. I can see serving such information directly, or embedding it into a larger JSON response.
2. Committed JSON files. This is an anti-pattern, but one I've seen many times. In one case, it was a translation file that was generated in a separate project. When I joined, it had been like that for more than 5 years. It was like that when I left, although I at least monkey patched REXML to emit sorted attributes in XML (also unordered in theory).
So while in practice, I don't depend on a specific ordering, but I need the order (whatever it is) to be stable. For all I care, it could be sorted by the cryptographic hash of the keys, so long as it is consistent.
Just to gather what others have said here: JSONB appears to store fields sorted lexicographically for binary search purposes, and emits them in that order as well. This could be undocumented behavior, yet important for the above two use cases.
Or you can sort the keys on serialization when you serve or store the JSON.
For hand-written JSON, I'm not sure there's a good solution besides modding the users' editors to sort for them.