[SPARK-47365][PYTHON] Add toArrow() DataFrame method to PySpark

[ianmcook]

What changes were proposed in this pull request?

• Add a PySpark DataFrame method toArrow() which returns the contents of the DataFrame as a PyArrow Table, for both local Spark and Spark Connect.
• Add a new entry to the Apache Arrow in PySpark user guide page describing usage of the toArrow() method.
• Add a new option to the method _collect_as_arrow() to provide more useful output when there are zero records returned. (This keeps the implementation of toArrow() simpler.)

Why are the changes needed?

In the Apache Arrow community, we hear from a lot of users who want to return the contents of a PySpark DataFrame as a PyArrow Table. Currently the only documented way to do this is to return the contents as a pandas DataFrame, then use PyArrow (pa) to convert that to a PyArrow Table.

pa.Table.from_pandas(df.toPandas())

But going through pandas adds significant overhead which is easily avoided since internally toPandas() already converts the contents of Spark DataFrame to Arrow format as an intermediate step when spark.sql.execution.arrow.pyspark.enabled is true.

Currently it is also possible to use the experimental _collect_as_arrow() method to return the contents of a PySpark DataFrame as a list of PyArrow RecordBatches. This PR adds a new non-experimental method toArrow() which returns the more user-friendly PyArrow Table object.

This PR also adds a new argument empty_list_if_zero_records to the experimental method _collect_as_arrow() to control what the method returns in the case when the result data has zero rows. If set to True (the default), the existing behavior is preserved, and the method returns an empty Python list. If set to False, the method returns returns a length-one list containing an empty Arrow RecordBatch which includes the schema. This is used by toArrow() which requires the schema even if the data has zero rows.

For Spark Connect, there is already a SparkSession.client.to_table() method that returns a PyArrow table. This PR uses that to expose toArrow() for Spark Connect.

Does this PR introduce any user-facing change?

• It adds a DataFrame method toArrow() to the PySpark SQL DataFrame API.
• It adds a new argument empty_list_if_zero_records to the experimental DataFrame method _collect_as_arrow() with a default value which preserves the method's existing behavior.
• It exposes toArrow() for Spark Connect, via the existing SparkSession.client.to_table() method.
• It does not introduce any other user-facing changes.

How was this patch tested?

This adds a new test and a new helper function for the test in pyspark/sql/tests/test_arrow.py.

Was this patch authored or co-authored using generative AI tooling?

No

[ianmcook]

cc @xinrong-meng@itholic@dongjoon-hyun

[HyukjinKwon]

Okay I'm fine with this

[ianmcook]

Thanks. I rebased. I also took a closer look at your changes in #46129 and made a few changes for consistency with the new structure you introduced there:

• I added a definition for toArrow in python/pyspark/sql/classic/dataframe.py.
• I added a placeholder definition for toArrow in python/pyspark/sql/dataframe.py
• I moved the docstring to there from python/pyspark/sql/pandas/conversion.py.

I also added .. versionadded:: 4.0.0 in the docstring.

Please let me know if that all looks OK.

[HyukjinKwon]

Merged to master.