Cudf: [BUG] rapidsai-nightly/label/cuda10.0 cudf.fillna(0) fail

[Polarbeargo]

Describe the bug
I use cudf 0.8 rapidsai-nightly/label/cuda10.0 as i implement
df.fillna(0) encouter
/databricks/python/lib/python3.6/site-packages/numba/cuda/compiler.py:234: UserWarning: Could not autotune, using default tpb of 128 warnings.warn('Could not autotune, using default tpb of 128') TypeError: fill_value must be a string or a string series

Steps/Code to reproduce bug

Expected behavior
Implement df.fillna(0) without error.
Environment details (please complete the following information):

[kkraus14]

It looks like you have some string dtype columns which can't just be filled with 0 naively. Pandas uses Python objects under the hood whereas we use an actual UTF8 string type.

[kkraus14]

@randerzander @beckernick any views on how this should best be handled from your perspective? Would you expect it to fill with "0" or throw an error as it currently does?

[randerzander]

IMO it's best to keep it an error as is.

This feels like one of those places where automatic type conversion (0 -> "0") is highly likely to induce confusion downstream.

Consider the scenario where CSV dtype inference reads a string when the user expected numerics. Prior to running some algorithm, they do:
for col in df.columns: df[col] = df[col].fillna(0)

At best, simple numeric operators surface the problem early.

At worst, a numba kernel or call to fit gives an inscrutable error, and significant time gets wasted trying to figure out what's going wrong.

[Polarbeargo]

If I write df.fillna("0") will bumped into following error message:

[kkraus14]

@polarbeargo trying to run fillna on the full dataframe object is going to be problematic for you since you have mixed dtypes. Here you're trying to fill numeric columns with a string which isn't allowed and we don't want to automatically typecast all of your numeric columns to strings.

[Polarbeargo]

Thank you @kkraus14 very appreciate your guidance:)!