gh-81022: Supporting customization of float encoding in JSON

[Lee-W]

Add an encode_float argument to JSONEncoder for supporting customization float encoding

• Issue: Supporting customization of float encoding in JSON #81022

[the-knights-who-say-ni]

Hello, and thanks for your contribution!

I'm a bot set up to make sure that the project can legally accept your contribution by verifying you have signed the PSF contributor agreement (CLA).

Our records indicate we have not received your CLA. For legal reasons we need you to sign this before we can look at your contribution. Please follow the steps outlined in the CPython devguide to rectify this issue.

If you have recently signed the CLA, please wait at least one business day
before our records are updated.

You can check yourself to see if the CLA has been received.

Thanks again for your contribution, we look forward to reviewing it!

[mitar]

I think also some tests would be useful?

[Lee-W]

@mitar I've added a test for the newly added encode_float argument.

[cmnord]

Rather than add another argument to these functions, why not un-nest floatstr from iterencode so that it can be overridden?

[mitar]

I think the idea is that one can also use C code-path for this? Why it is an argument for decoding and not a method?

[cmnord]

@mitar I don't fully understand your comment. Could you clarify what you mean?

What I meant is that it would be nice to be able to do the following:

importjsonimportnumpyasnpclassMyJSONEncoder(json.JSONEncoder):
deffloatstr(self, o, _repr=float.__repr__, _inf=json.INFINITY, _neginf=-json.INFINITY):
ifo!=o:
text="None"elifo==_inf:
text="Infinity"elifo==_neginf:
text="-Infinity"else:
return_repr(o)
ifnotself.allow_nan:
raiseValueError(
"Out of range float values are not JSON compliant: "+repr(o)
)
returntextassertjson.dumps(np.nan) =="None"

So rather than adding a new argument to many functions, we could achieve the same goal of custom float encoding this way. What do you mean by "C code-path" and "decoding and not a method"?

[mitar]

• In load we already have parse_float and parse_int. Why do you think this is like that, and not simply a method on cls?
• I think the reason why is because allows a fast code path in optimized C version of the parser.
• So, we should probably do a similar thing in encoding as well, no? So while your suggestion works, I think the idea is that it is faster if you just have to change those two. But we should probably measure that and see if truly is. But currently, the API here tries to match decoding. It might be confusing that for decoding you have function callbacks, but for encoding you have methods?

[cmnord]

@mitar thanks for clarifying, I think I understand now. 🙂

[mitar]

If you do have time, you could try to measure how much does this really benefit. You could try moving parsing of floats to a class in decoding. And see if that really slows down things. Personally I also do not like extra functions, especially if we already have a nice class to put methods on.

[Lee-W]

@mitar Could you please explain more on what should I measure? It seems you already explain it. Thanks 🙂

[mitar]

I would propose that you measure and compare two cases a) having methods on the class b) having functions directly provided.

[Lee-W]

I've tested on Python 3.8.0a4+ using the code below.
The result is as follows.

Encode
Default: 0.19853970199999998
Argument: 0.31803161799999996
Class: 0.282146192

importjsonimporttimeitINFINITY=json.encoder.INFINITYclassMyJSONEncoder(json.JSONEncoder):
deffloatstr(self, obj, _repr=float.__repr__, _inf=INFINITY, _neginf=-INFINITY):
ifobj!=obj:
text="None"elifobj==json.encoder.INFINITY:
text="Infinity"elifobj==-json.encoder.INFINITY:
text="-Infinity"else:
returnfloat.__repr__(obj)
ifnotself.allow_nan:
raiseValueError(
"Out of range float values are not JSON compliant: "+repr(obj)
)
returntextdeffloatstr(obj, allow_nan=True, _repr=float.__repr__, _inf=INFINITY, _neginf=-INFINITY):
ifobj!=obj:
text="None"elifobj==json.encoder.INFINITY:
text="Infinity"elifobj==-json.encoder.INFINITY:
text="-Infinity"else:
returnfloat.__repr__(obj)
ifnotallow_nan:
raiseValueError(
"Out of range float values are not JSON compliant: "+repr(obj)
)
returntextobj= {
'inf': float('inf'),
'-inf': -float('-inf'),
'nan': float('nan'),
'2': 2.23,
}
print('Encode')
print('Default: ', timeit.timeit(lambda: json.dumps(obj), number=10000))
print('Argument: ', timeit.timeit(lambda: json.dumps(obj, encode_float=floatstr), number=10000))
print('Class: ', timeit.timeit(lambda: json.dumps(obj, cls=MyJSONEncoder), number=10000))

[Lee-W]

@mitar Do I need to do other experiments on it? Or, would the above one be sufficient? 🙂

[mitar]

I will leave to somebody else from the Python team to way in here.

[mitar]

@Lee-W Please update the PR, there is now a merge conflict.

I really like this PR, could somebody from Python core team review/merge this?

[Lee-W]

@mitar Thanks for reminding. I just fix the conflict.

[merwok]

Minor comment: why move the floatstr function definition inline here? it could stay as a regular method in the class, then the line below would be self.encode_float = self.floatstr and still work.

[martin-remy]

@python's team : What's going on ?

[merwok]

As I noted on the ticket, this should be discussed to reach agreement on the need and the shape of the feature first.

[Lee-W]

Got it. I just wanted to resolve the previous conflict. As this is not yet agreed, let me close this PR