[FEA] Support nested column pruning in ORC reader when reading a struct column.

[firestarman]

Assuming there is an orc file containing one struct column "_c0" as below.

>>> pandas.read_orc("/data/tmp/test.log")
                                                  _c0
0                                                None
1   {'child0': -21.0, 'child1': 770290356.0, 'chil...
2   {'child0': 112.0, 'child1': 741485586.0, 'chil...
3   {'child0': -50.0, 'child1': -2094070070.0, 'ch...
4   {'child0': -49.0, 'child1': -228841867.0, 'chi...
5   {'child0': -55.0, 'child1': -2116626173.0, 'ch...
6   {'child0': None, 'child1': 2117211837.0, 'chil...
7   {'child0': -61.0, 'child1': -456120487.0, 'chi...
8                                                None
9   {'child0': 126.0, 'child1': -850819620.0, 'chi...
10  {'child0': 88.0, 'child1': 2143646677.0, 'chil...
11  {'child0': -18.0, 'child1': -1698102546.0, 'ch...
12  {'child0': -121.0, 'child1': 1460680657.0, 'ch...
>>> 

We can read only one nested column by Pandas orc reader, e.g

>>> pandas.read_orc("/data/tmp/test.log", columns=['_c0.child0'])
                   _c0
0                 None
1    {'child0': -21.0}
2    {'child0': 112.0}
3    {'child0': -50.0}
4    {'child0': -49.0}
5    {'child0': -55.0}
6     {'child0': None}
7    {'child0': -61.0}
8                 None
9    {'child0': 126.0}
10    {'child0': 88.0}
11   {'child0': -18.0}
12  {'child0': -121.0}
>>> 

or two of the nested columns,

>>> pandas.read_orc("/data/tmp/test.log", columns=['_c0.child1', '_c0.child0'])
                                           _c0
0                                         None
1     {'child0': -21.0, 'child1': 770290356.0}
2     {'child0': 112.0, 'child1': 741485586.0}
3   {'child0': -50.0, 'child1': -2094070070.0}
4    {'child0': -49.0, 'child1': -228841867.0}
5   {'child0': -55.0, 'child1': -2116626173.0}
6     {'child0': None, 'child1': 2117211837.0}
7    {'child0': -61.0, 'child1': -456120487.0}
8                                         None
9    {'child0': 126.0, 'child1': -850819620.0}
10    {'child0': 88.0, 'child1': 2143646677.0}
11  {'child0': -18.0, 'child1': -1698102546.0}
12  {'child0': -121.0, 'child1': 1460680657.0}

However cudf will complain an error for all the cases above.

>>> cudf.read_orc("/data/tmp/test.log", columns=['_c0.child0'])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/liangcail/miniconda3/envs/cudf-python-rt/lib/python3.7/site-packages/cudf/io/orc.py", line 302, in read_orc
    timestamp_type,
  File "cudf/_lib/orc.pyx", line 73, in cudf._lib.orc.read_orc
  File "cudf/_lib/orc.pyx", line 110, in cudf._lib.orc.read_orc
RuntimeError: cuDF failure at: ../src/io/orc/reader_impl.cu:540: Unknown column name : _c0.child0

Depends on #7830

[firestarman]

You can get the test file in issue #8704.

[devavret]

CC: @rgsl888prabhu

@vuule I think we also need to support this for parquet reader to address this #7248 (comment).

[firestarman]

One more thing.
The column order in the output of Pandas aligns with the file schema ('child0', 'child1'), even ['_c0.child1', '_c0.child0'] is specified.

I hope cuDF can support the output columns having the same order with the column names in parameter columns, meaning ('child1', 'child0'). Or at least having an addition parameter to control this behavior.

[devavret]

I hope cuDF can support the output columns having the same order with the column names in parameter columns, meaning ('child1', 'child0'). Or at least having an addition parameter to control this behavior.

Actually that's easier to support. It's what we do for non-nested column selection.