cumsum() changes dtype to object

[SGStino]

i

2019-12-28 00:00:52.470569300+00:00     -661212.7629936477
2019-12-28 00:00:56.358419200+00:00    -1383400.0425049507
2019-12-28 00:00:58.889932700+00:00     -894082.3358455619
2019-12-28 00:01:01.422784900+00:00     -874452.8101116775
                                              ...         
2019-12-28 23:57:23.719467400+00:00                    0.0
2019-12-28 23:58:00.080144500+00:00      2151.174441538055
2019-12-28 23:58:02.064671800+00:00                    0.0
2019-12-28 23:59:39.443425300+00:00      5761.132695684438
2019-12-28 23:59:41.927944600+00:00                    0.0
Length: 12928, dtype: pint[joule]

i.cumsum()

2019-12-28 00:00:52.470569300+00:00     -661212.7629936477 joule
2019-12-28 00:00:56.358419200+00:00    -2044612.8054985984 joule
2019-12-28 00:00:58.889932700+00:00    -2938695.1413441603 joule
2019-12-28 00:01:01.422784900+00:00     -3813147.951455838 joule
                                                 ...            
2019-12-28 23:57:23.719467400+00:00    -4767888585.5850525 joule
2019-12-28 23:58:00.080144500+00:00     -4767886434.410611 joule
2019-12-28 23:58:02.064671800+00:00     -4767886434.410611 joule
2019-12-28 23:59:39.443425300+00:00     -4767880673.277915 joule
2019-12-28 23:59:41.927944600+00:00     -4767880673.277915 joule
Length: 12928, dtype: object

this is likely caused by np.cumsum on an PintArray resulting in an numpy array of Quantity objects because it falls back to a loop of + operations?

My manual workaround is:
my_series.to_frame('power').pint.dequantify().cumsum().pint.quantify()

[andrewgsavage]

this is likely caused by np.cumsum on an PintArray resulting in an numpy array of Quantity objects because it falls back to a loop of + operations?
I think so. Pandas would receive a numpy array with nothing to tell it to interpret it as a PintArray so it uses object dtype.

I'm assuming this is the case for all numpy functions. I haven't looked how numpy operations work pandas, we may need to handle them within pint-pandas for them to work nicely.
This would use pint's numpy functionality hgrecco/pint#981, similar to what's done for addition/subtraction https://github.com/hgrecco/pint-pandas/blob/master/pint_pandas/pint_array.py#L539

For your workaround, there is a private method for creating a PintArray from a numpy array. I'm not sure what units your workaround results in.
PintArray._from_sequence( i.cumsum().values )

[MichaelTiemannOSC]

Just a note that I've run into this to. My workaround involves letting cumsum wreck the PintArray and then re-creating it with a call to astype(). But it would be so nice if cumsum would honor the Extension Array.

[andrewgsavage]

fixed by #186