Multicolumn GroupBy appears to convert unit64s to floats

[brianwgoldman]

Code Sample, a copy-pastable example if possible

pd.DataFrame({'first': [1], 'second': [1], 'value': [16148277970000000000]}).groupby(['first', 'second'])['value'].max()

Problem description

When that code snippet runs, the result is not 16148277970000000000 as you would expect, but 16148277969999998976. Note that int(float(16148277970000000000)) == 16148277969999998976.

Additional notes:

1. The problem only appears to happen if there are multiple groupby keys. For example just doing groupby(['first']) returns the expected result. So does removing the groupby statement entirely.
2. The problem is not specific to max. I get the same problem for min, first, last, median, mean, but not head, tail, or apply.
3. The problem also happens if you do .transform(max).
4. The smallest number I've found so far that has a problem is 2**63 + 1

Expected Output

16148277970000000000

Output of pd.show_versions()

INSTALLED VERSIONS

commit : None
python : 3.6.9.final.0
python-bits : 64
OS : Linux
OS-release : 4.14.137+
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_US.UTF-8
LOCALE : en_US.UTF-8

pandas : 0.25.3
numpy : 1.17.5
pytz : 2018.9
dateutil : 2.6.1
pip : 19.3.1
setuptools : 42.0.2
Cython : 0.29.14
pytest : 3.6.4
hypothesis : None
sphinx : 1.8.5
blosc : None
feather : 0.4.0
xlsxwriter : None
lxml.etree : 4.2.6
html5lib : 1.0.1
pymysql : None
psycopg2 : 2.7.6.1 (dt dec pq3 ext lo64)
jinja2 : 2.10.3
IPython : 5.5.0
pandas_datareader: 0.7.4
bs4 : 4.6.3
bottleneck : 1.3.1
fastparquet : None
gcsfs : 0.6.0
lxml.etree : 4.2.6
matplotlib : 3.1.2
numexpr : 2.7.1
odfpy : None
openpyxl : 2.5.9
pandas_gbq : 0.11.0
pyarrow : 0.14.1
pytables : None
s3fs : 0.4.0
scipy : 1.4.1
sqlalchemy : 1.3.12
tables : 3.4.4
xarray : 0.14.1
xlrd : 1.1.0
xlwt : 1.3.0
xlsxwriter : None

[WillAyd]

Can you try on master? Looks OK for me:

>>> pd.DataFrame({'first': [1], 'second': [1], 'value': [16148277970000000000]}).groupby(['first', 'second'])['value'].max()
first  second
1      1         16148277970000000000
Name: value, dtype: uint64

[brianwgoldman]

If you can reproduce the problem at 0.25.3 but not at master, I think we can call this closed as already fixed.

If not, I'm going to need your help on how to build the master version as I've not done it before.

[Dr-Irv]

Works in 1.1.1

[Dr-Irv]

A PR that creates a test case would be welcomed.

[TAJD]

take