[FEA] csv_reader_options to read empty strings as blank (i.e. ""), not null.

[mythrocks]

When cudf::io::read_csv() encounters two consecutive field delimiters within a row, it deems the corresponding string column value as null. E.g.:

a,,c
d,,f

Reading the input above via read_csv() produces rows {a,null,c} and {d,null,f}. This is conformant with Spark's CSV reader (and presumably Pandas).

It would be useful if the column value could be optionally interpreted as an empty string ("") instead. This would permit support for reading Hive delimited text, where empty strings are empty by default.

[davidwendt]

I believe you can post process this with cudf::replace_nulls with an empty string https://docs.rapids.ai/api/libcudf/stable/group__transformation__replace.html#gad359a898c2b11e70c3e33720259c5596

Technically you could also just remove the null mask. This assumes that null entries are created as empty strings with a corresponding validity bit set to 0.

[mythrocks]

you can post process this with cudf::replace_nulls with an empty string

The snag there is that we can't then differentiate between a legitimately null string and an empty one.
Consider that we have csv_reader_options.set_na_values({"\N"}) as is common in Spark/Hive. The second and third fields of the following input row cannot be represented via CUDF columns:

First,,\N,Last

The Hive format would treat the second field as empty (""), and the third as null.

[vuule]

This is surprising, I would expect set_na_values to impact this behavior. Checking if this is a bug.