Hi,

before 2.7, an
import json
json.loads('"abc"')

yielded u"abc".

in 2.7 I get
"abc" (a byte string).

I would have expected an entry in "news" or "What's new in 2.7" why this change happened. In addition, all examples at http://docs.python.org/library/json are wrong for Python 2.7  if json.loads is involved.

Any insight on this?

best regards,
Nik

This is related to this issue from simplejson:
http://code.google.com/p/simplejson/issues/detail?id=28

This problem is why I still use simplejson 1.x; moving forward to simplejson 2.x or Python's json is unlikely.

Well, then at least the documentation and the "What's changed" need to be updated. Furthermore, if such decisions are made, it would be at least nice to have some general "decode-hook" for json.JSONDecoder - the "object_hook" is only used for dict-objects - why is there no hook for strings or a general hook which is used on any objects?

As I understand it, the decision to return str instead of unicode
values for the "simplejson" module was simply inherited by the
standard library.  As such, it still needs to be evaluated in the
context of the standard library, because of the incompatibility it
introduces.

I still maintain that it's a bug, and should be treated as such.

Yep, the solution should not be "maybe it's str, maybe it's unicode" - I mean, if the decoder gives you a str if there are no fancy characters and unicode if it contains some, this might lead to some confusion... And yes, in my opinion, this is a bug, too.

I completely agree with Fred; this is a regression and a bug in Python 2.7 and should be fixed.  I have a doctest in Mailman 3 for example that cannot pass in both Python 2.6 and 2.7 (without IMO ugly hackery).  Not only that, but json is documented as converting JSON str to unicode, which it does fine in Python 2.6, 3.1 and 3.2.  Why should Python 2.7 be different (and broken)?

I'll note that it seems relevant that this package is not considered
"externally maintained" by the terms of PEP 360:

    http://www.python.org/dev/peps/pep-0360/

Given the level of attention this has received from the originator of
the code, we should not hesitate to commit technically acceptable
changes to the Python repository,

+1 for fixing this in-tree. We need a patch, though ;)

There is even more inconsistency here.

As already mentioned, we have this:


>>> import json
>>> json.loads(json.dumps("abc"))

'abc'

If, however, I am evil and hide _json.so (which is the C-part of the json module for speedup), the JSON code falls back to its python implementation and voila:

>>> import json
>>> json.loads(json.dumps("abc"))

u'abc'

Not so neat, if your fallback is not a fallback but shows such different behaviour.

Alternately, the Python implementation may be thought of as definitive
and the optimizations are broken.

The following patch (against the release27-maint branch) seems to fix the problem.

The fact that the C and Python versions are not fully tested (afaict) is not good.  I'm not sure that's worth fixing for 2.7 and it's probably worth a separate bug report for Python 3.2 on that.

In the meantime, I'll test Walter's patch and add a unit test for this case.

The incomplete testing and C/Python implementation mismatch are covered by issue 5723 and issue 9233.

To mitigate possible negative impacts from changing the return type, consider adding a parse_string hook that lets users control the return type:

   json.loads(f, parse_int=decimal.Decimal, parse_string=repr)

Adding that argument to Python 2.7 seems like new feature territory.

@doerwalter: patch looks good.  I've added a test and will commit momentarily.

r86126