Hyukjin Kwon created SPARK-61113:
------------------------------------
Summary: Add DataFrame.__arrow_c_stream__, the Arrow C stream of a
DataFrame, in the Python client as PySpark does
Key: SPARK-61113
URL: https://issues.apache.org/jira/browse/SPARK-61113
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: connect-rust-4.3.0
Reporter: Hyukjin Kwon
PySpark 4.3.0's DataFrame implements the stream of the Arrow PyCapsule
interface. {{DataFrame.__arrow_c_stream__(requested_schema=None)}} exports the
rows of the DataFrame as an Arrow C stream
({{pyspark.sql.interchange.SparkArrowCStreamer}}). It reads one partition at a
time, through {{mapInArrow}} and {{toLocalIterator}}, so the client never holds
all the rows at once. Libraries that read Arrow streams then take a Spark
DataFrame directly, for example {{pa.RecordBatchReader.from_stream(df)}},
Polars or DuckDB.
The Python client has neither {{DataFrame.__arrow_c_stream__}} nor
{{pyspark.sql.interchange}}. {{pa.RecordBatchReader.from_stream(df)}} raises
{{TypeError: Expected an object implementing the Arrow PyCapsule Protocol for
streams}}.
The fix is to vendor {{pyspark.sql.interchange}} from Apache Spark v4.3.0 and
to add {{__arrow_c_stream__}} to the DataFrame, as PySpark has them.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
