Hyukjin Kwon created SPARK-60999:
------------------------------------
Summary: Create the UDTFs of udtf and arrow_udtf of the Python
client as PySpark does
Key: SPARK-60999
URL: https://issues.apache.org/jira/browse/SPARK-60999
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: connect-rust-4.3.0
Reporter: Hyukjin Kwon
In PySpark Connect, {{arrow_udtf}} creates a PyArrow-native UDTF, of eval type
{{SQL_ARROW_UDTF}} (302), whose {{eval}} takes {{pyarrow}} arrays and returns
{{pyarrow}} tables. {{udtf}} creates one of {{SQL_TABLE_UDF}} (300), or of
{{SQL_ARROW_TABLE_UDF}} (301) when {{useArrow}} is True, or is not given and
{{spark.sql.execution.pythonUDTF.arrow.enabled}} of the session is true. Both
are not deterministic, unless {{asDeterministic()}} makes them so.
The Rust-backed Python client creates an {{arrow_udtf}} of
{{SQL_ARROW_TABLE_UDF}}, so the worker runs it as a Python UDTF whose {{eval}}
takes Python values, and every {{arrow_udtf}} fails. For example, PySpark
returns {{Row(x=2)}} for {{PlusOne(lit(1))}} of
{code:python}
@arrow_udtf(returnType="x long")
class PlusOne:
def eval(self, x):
yield pa.table({"x": pc.add(x, 1)})
{code}
where the client raises {{PYTHON_EXCEPTION}}. The client also ignores
{{spark.sql.execution.pythonUDTF.arrow.enabled}}, and creates deterministic
UDTFs. Create the UDTFs as PySpark does.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
