[BUG] can't merge dask-cudf dataframes when index is non-numeric?

[wence-]

Describe the bug

Consider

import cudf
import dask_cudf as dd

df1 = cudf.DataFrame({"a": ["a", "b"], "b": [1, 2]})
df2 = cudf.DataFrame({"a": ["a", "c"], "b": [2, 3]})

ddf1 = dd.from_cudf(df1, npartitions=2).set_index("a")
ddf2 = dd.from_cudf(df2, npartitions=2).set_index("a")

union = ddf1.merge(ddf2, left_index=True, right_index=True, how="left").compute()

This produces

File ~/cudf/python/cudf/cudf/core/column/string.py:5475, in StringColumn.as_numerical_column(self, dtype, **kwargs)
   5473 elif out_dtype.kind == "f":
   5474     if not libstrings.is_float(string_col).all():
-> 5475         raise ValueError(
   5476             "Could not convert strings to float "
   5477             "type due to presence of non-floating values."
   5478         )
   5480 result_col = _str_to_numeric_typecast_functions[out_dtype](string_col)
   5481 return result_col

ValueError: Could not convert strings to float type due to presence of
non-floating values.

This should work. I think something strange is going on, effectively
the error occurs because dask is trying to do

df3 = df1.set_index("a")

df4 = df3.assign(foo=cudf.Series([1, 2]))

And this fails because the assignment column doesn't have the same
index as the dataframe.

Whose bug is this? Everything "works" with pandas-backed dataframes.

[daxiongshu]

I'm curious about this bug. Thank you for raising the issue.

[wence-]

Looks like at least one problem is that .set_index() with a column of str dtype, divisions are not computed.

[rjzamora]

It isn't too difficult to get set_index to update the divisions correctly in dask_cudf. However, I'm still running into a separate issue where cudf doesn't support loc indexing of a range where one end of the range is not physicaly present in the index. For example:

import pandas as pd, cudf

pdf = pd.DataFrame({"a": ["a", "b", "c"], "b": range(3)}).set_index("a")
gdf = cudf.from_pandas(pdf)

pdf.loc["a":"d"]  # Works fine
gdf.loc["a":"d"]  # Fails

[rjzamora]

@wence- do you happen to know off-hand if this issue is already covered by something in #12793? If not, I can submit a new issue and link it.

[wence-]

I think it is not, please add to the list!

[wence-]

However, I'm still running into a separate issue where cudf doesn't support loc indexing of a range where one end of the range is not physicaly present in the index.

This is #12833.

[wence-]

It isn't too difficult to get set_index to update the divisions correctly in dask_cudf. However, I'm still running into a separate issue where cudf doesn't support loc indexing of a range where one end of the range is not physicaly present in the index. For example:

import pandas as pd, cudf

pdf = pd.DataFrame({"a": ["a", "b", "c"], "b": range(3)}).set_index("a")
gdf = cudf.from_pandas(pdf)

pdf.loc["a":"d"]  # Works fine
gdf.loc["a":"d"]  # Fails

This aspect should now be fixed. Can we revisit the original issue?

[wence-]

This looks like it is now fixed, we should probably add a test...