Are you using u(ltra)json over the standard library’s json module due to the performance benefits?

Until recently, I was too, but then I discovered that usjon.dumps() loses a lot of “precision” in floats compared to json.dumps():

>>> ujson.dumps(1.3959197885670265)
'1.3959197886'

>>> json.dumps(1.3959197885670265)
'1.3959197885670265'

You might now be thinking to yourself, doesn’t usjon.dumps() take a double_precisionparameter that allows you to control the “maximum digit precision of doubles”?

Well, you would be correct. Let’s try it:

>>> ujson.dumps(1.3959197885670265, double_precision=20)
'1.395919788567026'

>>> json.dumps(1.3959197885670265)
'1.3959197885670265'

OK, so '1.395919788567026' is definitely more “precise” than ‘1.3959197886’ which was returned by the call without the double_precision argument. However, if you look closely you’ll notice that even with double_precision=20 there’s still a single place of precision lost by ujson.dumps() compared to json.dumps(). This is despite the fact that setting double_precision to 20 should be more than enough to retain the full precision of the float in the example.

At the time of writing, I can’t find a way to force ujson.dumps() to use arbitrary float precision. Because of this, I’ve switched back to plain old json.dumps() in the application where I originally discovered this issue, because not losing any precision is more important than performance.