BUG: conversion of uint64 to float when taking

[jreback]

62fe71e#r100575113

# master
In [1]: pandas.core.algorithms.take_nd(np.array([0, 1], dtype='uint64'), np.array([0, -1]))
Out[1]: array([  0.,  nan])

# with this change above
In [5]: pandas.core.algorithms.take_nd(np.array([0, 1], dtype='uint64'), np.array([0, -1]))
Out[5]: array([                   0, 18446744073709551615], dtype=uint64)

mainly need some more testing w.r.t. this (e.g. pandas.types.cast._maybe_promote).

[jreback]

cc @gfyoung

[jreback]

this is not going to work as lots of code relies on this. This is using the np.iinfo(np.uint64).max as the missing value value.

In any event we should do something here, maybe treat uint with a missing value as an object dtype for now.

[jreback]

I took a different path in #15245

[jbrockmendel]

im not clear on what the bug is here. the behavior looks correct (both in OP and master)

[mroeschke]

Agreed, closing as this looks like the correct behavior