Int64Dtype in read_csv leads to unexpected values

[alohr]

Code Sample, a copy-pastable example if possible

import pandas as pd
import io

t = io.StringIO('''\
event,timestamp
a,1556559573141592653
b,1556559573141592654
c,
d,1556559573141592655
''')

# Reading the timestamps as strings works fine
print("\nExpected output:")
print(pd.read_csv(t, dtype={'timestamp': object}))

# Now with Int64Dtype
t.seek(0)
print("\nActual output:")
print(pd.read_csv(t, dtype={'timestamp': pd.Int64Dtype()}))

Problem description

I would like to read csv files with nullable (big) integers into a dataframe. The integers represent nanoseconds since the UNIX epoch 1970. Using the Int64Dtype introduced in 0.24.0 seems like the way to go. I quote from the FAQ:

https://pandas.pydata.org/pandas-docs/stable/user_guide/gotchas.html#nan-integer-na-values-and-na-type-promotions

If you need to represent integers with possibly missing values, use one of the nullable- integer extension dtypes provided by pandas

Expected Output

  event            timestamp
0     a  1556559573141592653
1     b  1556559573141592654
2     c                  NaN
3     d  1556559573141592655

Actual Output

  event            timestamp
0     a  1556559573141592576
1     b  1556559573141592576
2     c                  NaN
3     d  1556559573141592576

Output of pd.show_versions()

INSTALLED VERSIONS ------------------ commit: None

pandas: 0.24.2
pytest: None
pip: 19.1
setuptools: 41.0.1
Cython: None
numpy: 1.16.3
scipy: None
pyarrow: None
xarray: None
IPython: None
sphinx: None
patsy: None
dateutil: 2.8.0
pytz: 2019.1
blosc: None
bottleneck: None
tables: None
numexpr: None
feather: None
matplotlib: None
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml.etree: None
bs4: None
html5lib: None
sqlalchemy: None
pymysql: None
psycopg2: None
jinja2: None
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None
gcsfs: None

[WillAyd]

Yea I think there may be some undesired casting to float64 going on behind the scenes. Looks OK up to 2^52 but beyond that things get dicey:

In [15]: 2**53
Out[15]: 9007199254740992

t = io.StringIO('''\
event,timestamp
a,9007199254740992
b,9007199254740993
c,
d,9007199254740994
''')

In [20]: print(pd.read_csv(t, dtype={'timestamp': pd.Int64Dtype()}))
  event         timestamp
0     a  9007199254740992
1     b  9007199254740992
2     c               NaN
3     d  9007199254740994

Investigation and PRs are certainly welcome

[phofl]

Duplicate of #32134