[jira] [Created] (ARROW-7628) PyArrow read_csv problematic cases

["Athanassios Hatzis (Jira)" <ji...@apache.org>]

Athanassios Hatzis created ARROW-7628:
-----------------------------------------

             Summary: PyArrow read_csv problematic cases
                 Key: ARROW-7628
                 URL: https://issues.apache.org/jira/browse/ARROW-7628
             Project: Apache Arrow
          Issue Type: Bug
          Components: Python
    Affects Versions: 0.15.1
         Environment: Ubuntu bionic
            Reporter: Athanassios Hatzis
         Attachments: spc_catalog.tsv

Hi, I have found two problematic cases, possibly bugs in pyarrow read_csv module. I have written the following piece of code and run a test on the attached CSV file. 

The code compares pandas read_csv with pyarrow csv to show that the second is not behaving correctly with the following set of parameters:

1. change parameter skip_rows = 10, 
{code:python}
Traceback (most recent call last):
  File "/home/athan/anaconda3/envs/TRIADB/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-21-8c5c88b190c4>", line 4, in <module>
    read_options=csv.ReadOptions(skip_rows=skip_rows, autogenerate_column_names=False, use_threads=True, column_names=column_names)
  File "pyarrow/_csv.pyx", line 541, in pyarrow._csv.read_csv
  File "pyarrow/error.pxi", line 84, in pyarrow.lib.check_status
pyarrow.lib.ArrowKeyError: Column 'catcost' in include_columns does not exist in CSV file
{code}

2. skip_rows = 12, columns = None
In this case you don't get the error above, projection is None, but compare the two dataframes, the one from pyarrow with to_pandas() and the one from the output of pandas read_csv(). You will notice that the first one has not parsed correctly the null values ('\\N') in the last column catname. On the contrary pandas read_csv managed to parse all the null values correctly.

{code:python}
Out[28]: 
   1082  991   16.5    200 2014-09-10  1  bar
0  1082  997   0.55  100.0 2014-09-10  1  bar
1  1082  998   7.95  200.0 2014-03-03  0   \N
2  1083  998  12.50    NaN        NaT  0  bar
3  1083  999   1.00    NaN        NaT  0  foo
4  1084  994  57.30  100.0 2014-12-20  1   \N
5  1084  995  22.20    NaN        NaT  0  foo
6  1084  998  48.60  200.0 2014-12-20  1  foo

{code}

Python code to test the attached CSV file for the bugs reported above


{code:python}
from pyarrow import csv
import pyarrow as pa
import pandas as pd

file_location = 'spc_catalog.tsv'

sep = '\t'
nulls=['\\N']

columns = ['catcost', 'catqnt', 'catdate', 'catchk', 'catname']
column_names = None
column_types = None

skip_rows = None
nrecords = None

csv.read_csv(file_location,
    parse_options=csv.ParseOptions(delimiter=sep),
    convert_options=csv.ConvertOptions(include_columns=columns, column_types=column_types, null_values=nulls),
    read_options=csv.ReadOptions(skip_rows=skip_rows, autogenerate_column_names=False, use_threads=True, column_names=column_names)
).to_pandas()

pd.read_csv(file_location, sep=sep, na_values='\\N', usecols=columns, nrows=nrecords, names=column_names, dtype=column_types)

{code}



--
This message was sent by Atlassian Jira
(v8.3.4#803005)