Yicong Huang created SPARK-55224:
------------------------------------
Summary: Use Spark DataType as ground truth in Pandas-Arrow
serialization
Key: SPARK-55224
URL: https://issues.apache.org/jira/browse/SPARK-55224
Project: Spark
Issue Type: Improvement
Components: PySpark
Affects Versions: 4.2.0
Reporter: Yicong Huang
Currently, PySpark's Pandas serializers pass both {{spark_type}} and
{{arrow_type}} through the serialization pipeline. This creates confusion about
which type is authoritative. Analysis shows {{spark_type}} (from user's UDF
return type) is always the source, and {{arrow_type}} is derived via
{{to_arrow_type()}} which is lightweight.
This ticket proposes refactoring {{_create_batch}} and {{_create_array}} to
take only {{spark_type}} and derive {{arrow_type}} internally.
{code:python}
# Before: caller pre-computes arrow_type
arrow_return_type = to_arrow_type(return_type, ...)
yield (result, arrow_return_type, return_type)
# After: only pass spark_type
yield (result, return_type)
{code}
{code:python}
# Before: _create_array takes arrow_type as primary
def _create_array(self, series, arrow_type, spark_type=None, arrow_cast=False):
dt = spark_type or from_arrow_type(arrow_type) # confusing fallback
# After: spark_type is ground truth
def _create_array(self, series, spark_type, *, arrow_cast=False,
prefers_large_types=False):
arrow_type = to_arrow_type(spark_type, ...) if spark_type else None
{code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
