To fix issues with floating-point precision we've made use of Google's
double-conversion lib to handle conversions of doubles to and from strings.
In addition to fixing our precision problems this will improve double
encoding by 4-5x. Decoding is however slightly slower according to the
benchmarks - but accurate at least.
This change removes the double_precision encoding option and the
precise_float decoding option.
To better align with the standard json module this removes ujson
default serialization of date/datetime objects to unix-timestamps.
Trying to serialize such an object will now raise a TypeError "repr(obj)
is not JSON serializable".
The behavior of ujson has always been to try to serialize all objects in
any way possible. This has been quite a deviation from other json
libraries, including Pythons standard json module, and the source of a
lot of confusion and bugs. Removing this quirk moves ultrajson closer to
the expected behavior.
Instead of trying to coerce serialization ultrajson will now throw a
TypeError: "repr(obj) is not JSON serializable" exception.
Previously a None dict item key would be outputted in JSON as "None".
To better align with the standard json module this was changed to output
"null". There's no proper representation of null object keys in JSON so
this is implementation specific but it seems more natural to follow
suit when it can be done without a significant performance hit.
Added and used branch prediction macros (LIKELY/UNLIKELY) as well.
This was caused by checking for "__json__" using PyObject_HasAttrString
which clears the error set by a previous long overflow. Thus this was dependent
on the order of processing of dict items, which explains why it was
seemingly random as the dict items are likely ordered by a hash of
the key.
This fixes GH224 and GH240.
- Fixed segfault when using sort_keys=True on dict with unorderable keys (GH247)
- Fixed refcount becoming negative when using sort_keys=True (GH243)
- Fixed compile error when defining JSON_NO_EXTRA_WHITESPACE
caused by a wrongly named variable. (GH245)
1. It reduces clutter in symbol table.
2. It fixes issues with C99 inline semantics for functions
marked as inline (#237, #180, #222), which manifests
when compiled with GCC>=5.
this fixes issue #179.
setuptools should itself know when to use cache or create a
new build... however if someone wants to override that, it's
still possible but forcefully doing that on whatever
setuptools target will (and does) introduce problems.
Build directory should be cleaned up via the clean sub-command.
examples:
- clean up temp:
python setup.py clean
- clean up whole build dir
python setup.py clean -a
Or if somebody wants to, the build dir could be removed on the
shell.
both Python 2 and Python 3, removing the need for 2to3 to be run. This
indirectly fixes #177.
* Corrected a duplicate-named method in the unit testing code. Corrected the
now-exposed broken logic that wasn't being tested. This was
highlighted in #186 but the author did not appear to realise that the
method name was being masked in the unit test class.
* Corrected the test case skipping logic to use unittest.skipIf instead
of just returning from the method upon an ImportError so that the skip
can actually be registered as a skipped test instead of a passed test.
* Updated the tests to additionally run on Python 3.5 on Travis. This
covers #195.
* Merged the two benchmarking files into one, and modularised the code.
Also added native RST output so the benchmarking results can be placed
directly into the README file.