[BUG][GPU Error Bug] "SELECT ((<string>)!=(<column>)) FROM <tables>" brings Error about dask-sql HOT 5 OPEN

I have also encountered the same situation. Can the developer reply？

from dask-sql.

It looks like the underlying issue here is that dask-cuDF is failing because we're trying to do a binop between a float column and a string scalar:

import cudf
import dask_cudf

s = cudf.Series(["A"])

ds = dask_cudf.from_cudf(s, npartitions=1)

ds != "A"

On dask-sql's end, it seems like the specific combination of cross joins planned when selecting from all 3 tables in your reproducer is making it such that t1_gpu.c0 is registered a float column rather than the string/object column I would've expected - will dig into this deeper to get a sense of what's happening

from dask-sql.

Okay seems like the issue here is handling of duplicate column names on GPU, it looks like the cross join itself fails with 3 tables:

import pandas as pd
import dask.dataframe as dd
from dask_sql import Context

c = Context()
c.create_table('df1', pd.DataFrame({"a": [1]}), gpu=True)
c.create_table('df2', pd.DataFrame({"a": [2]}), gpu=True)
c.create_table('df3', pd.DataFrame({"a": [3]}), gpu=True)

query = "SELECT * FROM df1, df2, df3"
explain = c.explain(query)
res = c.sql(query)  # AssertionError

There's a good chance that this could be related to #1133 and potentially resolved with #1134, I can look into reviving that PR and giving it a try here

EDIT:

Tried #1134 without much luck, looks like the cross join issue is independent of CPU/GPU as I get errors with the above block even when gpu=False, think it just ends up failing on GPU for your particular reproducer because dask-cuDF doesn't support float column / string scalar binops whereas Dask CPU does

from dask-sql.

Trying out your reproducer with #1250, I'm now able to get things passing:

CPU Result:
   Utf8("A") != t1.c0
0                True
GPU Result:
   Utf8("A") != t1_gpu.c0
0                    True

Would you mind giving this branch a look on your end?

from dask-sql.

The bug came up at dask-sql version: 2023.6.0. After my verification, the bug was fixed at #1250 .
Thanks for your work.

from dask-sql.