[
https://issues.apache.org/jira/browse/ARROW-17912?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17612897#comment-17612897
]
Joris Van den Bossche commented on ARROW-17912:
-----------------------------------------------
I agree with David that it seems this can be solved in pyspark.
You referenced
https://github.com/apache/spark/blob/5483607910aba0aaf05d029c6c813faa37cb4731/python/pyspark/sql/pandas/serializers.py#L366
where {{Table.from_batches}} is being called, which requires a schema if
{{batches}} is an empty list. But those batches comes from
{{ArrowStreamSerializer.load_stream}}, which calls {{pyarrow.ipc.open_stream}}
(https://github.com/apache/spark/blob/5483607910aba0aaf05d029c6c813faa37cb4731/python/pyspark/sql/pandas/serializers.py#L95-L97).
At that point, you always have the schema from the stream as well
({{reader.schema}}), so it is just a matter of passing this through to
{{from_batches}}.
> [C++] IPC writer does not write an empty batch in the case of an empty table,
> which PySpark cannot handle
> ---------------------------------------------------------------------------------------------------------
>
> Key: ARROW-17912
> URL: https://issues.apache.org/jira/browse/ARROW-17912
> Project: Apache Arrow
> Issue Type: Bug
> Components: C++
> Reporter: Liangcai li
> Priority: Major
>
> My current work is about Pyspark Cogroup Pandas UDF. And two processes are
> involved, the JVM one (sender) and the Python one (receiver).
> [Spark is using the Arrow Java
> `ArrowStreamWriter`|https://github.com/apache/spark/blob/branch-3.3/sql/core/src/main/scala/org/apache/spark/sql/execution/python/CoGroupedArrowPythonRunner.scala#L99]
> to serialize Arrow tables being sent from the JVM process to the Python
> process, and ArrowStreamWriter can handle empty tables correctly.
> [While cuDF is using the Arrow C++ RecordBatchWriter
> |https://github.com/rapidsai/cudf/blob/branch-22.10/java/src/main/native/src/TableJni.cpp#L254]to
> do the same serialization, but it leads to an error as below on the Python
> side, where [the Pyspark is calling Pyarrow
> *Table.from_batches*|https://github.com/apache/spark/blob/branch-3.3/python/pyspark/sql/pandas/serializers.py#L366]
> to deserialize the arrow stream.
> ```
> _E File
> "/usr/local/spark/python/lib/pyspark.zip/pyspark/sql/pandas/serializers.py",
> line 297, in load_stream_
> _E [self.arrow_to_pandas(c) for c in
> pa.Table.from_batches(batch2).itercolumns()]_
> _E File "pyarrow/table.pxi", line 1609, in
> pyarrow.lib.Table.from_batches_
> _E {color:#de350b}*ValueError: Must pass schema, or at
> least one RecordBatch*{color}_
> ```
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
