Timedelta to_json with date_format='iso'

[TomAugspurger]

Is this formatting for timdelta columns intended / correct?

In [12]: df = pd.DataFrame({"A": pd.date_range("2016", periods=4, freq='D'),
    ...:                    "B": pd.timedelta_range('1D', periods=4, freq='D')})

In [13]: df.to_json(date_format='iso')
Out[13]: '{"A":{"0":"2016-01-01T00:00:00.000Z","1":"2016-01-02T00:00:00.000Z","2":"2016-01-03T00:00:00.000Z","3":"2016-01-04T00:00:00.000Z"},"B":{"0":"1970-01-02T00:00:00.000Z","1":"1970-01-03T00:00:00.000Z","2":"1970-01-04T00:00:00.000Z","3":"1970-01-05T00:00:00.000Z"}}'

So a pd.Timedelta('1D') goes to "1970-01-02T00:00:00.000Z".

The expected might be a duration like

"P1DT0H0M0S".

It seems to be reusing the logic for datetime columns and treating it as the equivalent int -> iso format. Ideally to_json will have a new paramater timedelta_format that takes the same values as date_format.

[TomAugspurger]

xref #15136 and #14904

I can work on this, if we determine that it should be changed.

[jreback]

yes this is probably unsupported you can pass a default_handler to deal with it, though prob better do it in the c-code.

[TomAugspurger]

I'm gonna attempt to do it in objToJSON.c, we'll see how that goes 😆

[jreback]

yeah looks like its the iso code :>

[ron819]

@TomAugspurger any progress with this?

[jreback]

@ron819 why don’t you do a pull request for this? we have thousands of issues and very few core devs, who mostly do code review

[mroeschke]

Looks like this is solved on master by by: #30903