np.array([pd.Series({'a':'a'})]) works but np.array([pd.Series({1:1})]) doesn't

[soerenwolfers]

The last line here throws a KeyError: 0

import numpy as np
import pandas as pd
np.array([pd.Series({0:0})])
np.array([pd.Series({'a':'a'})])
np.array([pd.Series({1:1})])

Is that a bug?

[jorisvandenbossche]

Can you show the output of pd.show_versions() ?

[MarcoGorelli]

I can't reproduce this on v1.1.5

[soerenwolfers]

Sorry,

pandas : 1.1.4
numpy : 1.17.4

[MarcoGorelli]

Thanks, can confirm this reproduces.

If you upgrade pandas to 1.1.5 you'll find it fixed - it would be good to do a git bisect to figure out when this was fixed, and (if it was fixed accidentally) to add a test

[simonjayhawkins]

can't reproduce with 0.25.3, 1.0.5, 1.1.3, 1.1.4, 1.1.5, master (all envs using numpy 1.19+)

[MarcoGorelli]

Same, just tried in 1.1.4 with numpy 1.19.4 and it works fine.

Conversely, on master and with numpy 1.17.4, it still fails. I think we can close this as an upstream bug anyway

[jorisvandenbossche]

But is it an identified / reported upstream issue?
I think it's still good to add a test (it's not given that this was an intentional/tested fix on numpy's side)

[MarcoGorelli]

Sure - @soerenwolfers do you feel like contributing a test for this issue? If so, please check the contributing guide

[seberg]

Wow, there was an additional layer of confusion to unravel here :) (for myself). If there is only a single item in the dataframe, NumPy used to do use series[0] to get that item. But if there were more items it effectively used list(series). Since list() probably uses __iter__ internally – and that goes by position and not index – the only thing that gives different results on new NumPy seems to be the specific case of a series with a single element and non 0 index.

[horaceklai]

take

[horaceklai]

Hi I'm not sure what to do now. I made a PR and did the changes requested, but when running the tests, it throws the Key error: 0 as stated in the beginning of this issue.

#42939