[Python] parquet.read_table nested fields in columns does not work for use_legacy_dataset=False

[asfimport]

Reading nested field does not work with use_legacy_dataset=False.

This works:

 

import pyarrow.parquet as pq
t = pq.read_table(
 source=*filename*,
 columns=['store_key', 'properties.country'], 
 use_legacy_dataset=True,
).to_pandas()

This does not work (for the same parquet file):

 

import pyarrow.parquet as pq

t = pq.read_table(
 source=*filename*,
 columns=['store_key', 'properties.country'], 
 use_legacy_dataset=False,
).to_pandas()

 

Reporter: Tom Scheffers
Assignee: Miles Granger / @milesgranger

Related issues:

• [Python] Allow to create field reference to nested field (is blocked by)
• [Python] Can not refer to field in a list of structs (is related to)

PRs and other links:

• GitHub Pull Request #14326

_{Note: This issue was originally created as ARROW-14596. Please see the migration documentation for further details.}

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
[~TomScheffers] that's indeed a current limitation of the new implementation. There is work underway to enable this (the basic feature was just merged for C++ (ARROW-13987), now it still has to be exposed in python (ARROW-11259)), and hopefully this will be possible in the next version 7.0.0

[asfimport]

Krisztian Szucs / @kszucs:
Postponing to 9.0

[asfimport]

Alenka Frim / @AlenkaF:
I would like to add observations we got today when pairing with @jorisvandenbossche on this topic.

First was the result of using pq.read_table with legacy implementation vs using ds.dataset with column projection. The data can get selected correctly with the dataset implementation but what happens is that the structure of a nested field is not kept (from struct it is flattened to string column). In case of using columns selection with a list in  {}ds.dataset{}, it errors, as reported in the issue.

>>> import pandas as pd
>>> import pyarrow as pa
>>> import pyarrow.parquet as pq
>>> 
>>> df = pd.DataFrame({
...     'user_id': ['abc123', 'qrs456'],
...     'interaction': [{'type': 'click', 'element': 'button'}, {'type':'scroll', 'element': 'window'}]
... })
>>> 
>>> table = pa.Table.from_pandas(df)
>>> pq.write_table(table, 'example.parquet')

>>> pq.read_table('example.parquet', columns = ['user_id', 'interaction.type'], use_legacy_dataset = True)
pyarrow.Table
user_id: string
interaction: struct<type: string>
  child 0, type: string
----
user_id: [["abc123","qrs456"]]
interaction: [
  -- is_valid: all not null
  -- child 0 type: string
["click","scroll"]]

>>> import pyarrow.dataset as ds
>>> projection = {
...         'user_id': ds.field('user_id'),
...         'new': ds.field(('interaction', 'type'))
...     }
>>> ds.dataset('example.parquet').to_table(columns=projection)
pyarrow.Table
user_id: string
new: string
----
user_id: [["abc123","qrs456"]]
new: [["click","scroll"]]

>>> ds.dataset('example.parquet').to_table(columns=['user_id', 'interaction.type'])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "pyarrow/_dataset.pyx", line 303, in pyarrow._dataset.Dataset.to_table
    return self.scanner(**kwargs).to_table()
  File "pyarrow/_dataset.pyx", line 270, in pyarrow._dataset.Dataset.scanner
    return Scanner.from_dataset(self, **kwargs)
  File "pyarrow/_dataset.pyx", line 2322, in pyarrow._dataset.Scanner.from_dataset
    _populate_builder(builder, columns=columns, filter=filter,
  File "pyarrow/_dataset.pyx", line 2168, in pyarrow._dataset._populate_builder
    check_status(builder.ProjectColumns([tobytes(c) for c in columns]))
  File "pyarrow/error.pxi", line 100, in pyarrow.lib.check_status
    raise ArrowInvalid(message)
pyarrow.lib.ArrowInvalid: No match for FieldRef.Name(interaction.type) in user_id: string
interaction: struct<element: string, type: string>
__fragment_index: int32
__batch_index: int32
__last_in_fragment: bool
__filename: string
/Users/alenkafrim/repos/arrow/cpp/src/arrow/type.h:1722  CheckNonEmpty(matches, root)
/Users/alenkafrim/repos/arrow/cpp/src/arrow/type.h:1757  FindOne(root)
/Users/alenkafrim/repos/arrow/cpp/src/arrow/dataset/scanner.cc:714  ref->GetOne(dataset_schema)
/Users/alenkafrim/repos/arrow/cpp/src/arrow/dataset/scanner.cc:784  ProjectionDescr::FromNames(std::move(columns), *scan_options_->dataset_schema)

When Scanner object is being created from the dataset class via to_table and (through _populate_builder) and in the case of a list of columns the ProjectColumns method ("arrow::dataset::ScannerBuilder") is being called it only accepts string column names and errors when a column is a struct.

We were thinking if it would be a good idea to add a new method in scanner.cc that would mimic FromNames method but takes field_ref as an argument? Afterwords there would also be a need to recreate a struct field for which we are not sure how to approach.

cc @westonpace @pitrou do you think that would be a correct way to go?

[igozali]

Any updates for fixing this issue by any chance?