[FEA] perform shift operator with string column

[rnyak]

Is your feature request related to a problem? Please describe.

I am getting the following error from my code (see below):

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<timed exec> in <module>

/opt/conda/lib/python3.8/site-packages/cudf/core/frame.py in shift(self, periods, freq, axis, fill_value)
   1640         """
   1641         assert axis in (None, 0) and freq is None
-> 1642         return self._shift(periods)
   1643 
   1644     def _shift(self, offset, fill_value=None):

/opt/conda/lib/python3.8/site-packages/cudf/core/frame.py in _shift(self, offset, fill_value)
   1645         data_columns = (col.shift(offset, fill_value) for col in self._columns)
   1646         data = zip(self._column_names, data_columns)
-> 1647         return self.__class__._from_table(Frame(data, self._index))
   1648 
   1649     def __array__(self, dtype=None):

cudf/_lib/table.pyx in cudf._lib.table.Table.__init__()

/opt/conda/lib/python3.8/site-packages/cudf/core/column_accessor.py in __init__(self, data, multiindex, level_names)
    119             self._data = {}
    120             if data:
--> 121                 data = dict(data)
    122                 # Faster than next(iter(data.values()))
    123                 column_length = len(data[next(iter(data))])

/opt/conda/lib/python3.8/site-packages/cudf/core/frame.py in <genexpr>(.0)
   1643 
   1644     def _shift(self, offset, fill_value=None):
-> 1645         data_columns = (col.shift(offset, fill_value) for col in self._columns)
   1646         data = zip(self._column_names, data_columns)
   1647         return self.__class__._from_table(Frame(data, self._index))

/opt/conda/lib/python3.8/site-packages/cudf/core/column/column.py in shift(self, offset, fill_value)
    386 
    387     def shift(self, offset: int, fill_value: ScalarLike) -> ColumnBase:
--> 388         return libcudf.copying.shift(self, offset, fill_value)
    389 
    390     @property

cudf/_lib/copying.pyx in cudf._lib.copying.shift()

RuntimeError: cuDF failure at: /workspace/build-env/cpp/src/copying/shift.cu:51: shift does not support non-fixed-width types.

Describe the solution you'd like

You can reproduce the error with the following code:

df = cudf.read_parquet('df_toy.parquet')

# Keeps repeated interactions on the same items, removing only consecutive interactions, 
df = df.sort_values(['user_session', 'event_time_ts']).reset_index(drop=True)

print("Count with in-session repeated interactions: {}".format(len(df)))
# Sorts the dataframe by session and timestamp, to remove consecutive repetitions
df['product_id_past'] = df['product_id'].shift(1).fillna(0)
df['session_id_past'] = df['user_session'].shift(1).fillna(0)
#Keeping only no consecutive repeated in session interactions
df = df[~((df['user_session'] == df['session_id_past']) & \
             (df['product_id'] == df['product_id_past']))]
print("Count after removed in-session repeated interactions: {}".format(len(df)))
del(df['product_id_past'])
del(df['session_id_past'])

Dataset can be downloaded from here.

Describe alternatives you've considered

As a workaround, I have to label encode the user_session column first to be able to perform shift operation.
In addition, I can run this pipeline with pandas with long strings.

[davidwendt]

Strings column support was added to cudf::shift() in release 21.08 in PR #8648
The runtime error included in the description

RuntimeError: cuDF failure at: /workspace/build-env/cpp/src/copying/shift.cu:51: shift does not support non-fixed-width types.

indicates this was tested on a build prior to 21.08.

[beckernick]

If you run into any issues with shift after upgrading versions, please feel free to file an issue. For now, I'm going to close this issue as answered.