ENH: support reductions for pyarrow temporal types

[jbrockmendel]

We could plausibly disallow mean for date types, matching what we do for PeriodDtype.

Fixes 232 xfails, cuts test runtime by 88 seconds.

[mroeschke]

Is this on an older version of pyarrow?

In [6]: t = pa.time32("ms")

In [7]: t.unit
Out[7]: 'ms'

In [8]: pa.__version__
Out[8]: '10.0.1'

In [9]: t = pa.time64("ns")

In [10]: t.unit
Out[10]: 'ns'

In [11]: t = pa.time32("ms")

In [12]: t.unit
Out[12]: 'ms'

[jbrockmendel]

looks like it depends on exactly what we're checking. Here pa_type is defined as pa_type = self._data.type which DataType(time32[ms]), which does not have a .unit attribute in 10.0.1

[jbrockmendel]

so how do I get the Time32Type(time32[ms]) instead of the DataType(time32[ms]) here?

[mroeschke]

Ah I see. Yeah I don't see a way to do that. Any insight @jorisvandenbossche?

[jorisvandenbossche]

Yeah, this was only recently fixed (should be in pyarrow 11): apache/arrow#14633

When getting the type from the array, it wasn't correctly boxed in the Time specific class, and so the unit attribute is not available. I am not directly sure of a workaround for older pyarrow versions, except for parsing that from the str repr of the type ...

[jbrockmendel]

so something like?

def get_unit_from_pyarrow_dtype(pa_dtype):
    if pa_lt_11:
        return str(pa_dtype).split("[")[-1][:-1]
    return pa_dtype.unit

[jorisvandenbossche]

Yes, that should work I think

[mroeschke]

Thanks @jbrockmendel