Hyukjin Kwon created SPARK-61846:
------------------------------------
Summary: Export is_remote_only from pyspark, make the serializers
and both ResourceProfile classes PySpark's, and give the modules the names that
PySpark's doctests use
Key: SPARK-61846
URL: https://issues.apache.org/jira/browse/SPARK-61846
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: connect-rust-4.4.0
Reporter: Hyukjin Kwon
Close small gaps of the Python drop-in against pyspark-client 4.3.0:
- "from pyspark import is_remote_only" raises ImportError; upstream's
pyspark/__init__.py imports it (and sys).
- pyspark.serializers.SpecialLengths lacks START_OF_INIT_MESSAGE (-8), which
the vendored pyspark.messages.socket.spark_socket_message_receiver reads.
serializers.py is otherwise upstream's file.
- pyspark.sql.connect.resource.profile.ResourceProfile is the classic
pyspark.resource.ResourceProfile, and both take (_exec_req=None,
_task_req=None). PySpark has two classes:
pyspark.resource.ResourceProfile(_java_resource_profile=None, _exec_req=None,
_task_req=None) and the connect class ResourceProfile(exec_req=None,
task_req=None), which creates the profile on the server through
SparkSession.client._create_profile.
- The official doctests of the Spark Connect modules take a module's namespace
as their globals, and need names there that the drop-in's modules lack: Row,
StructType and StorageLevel in pyspark.sql.dataframe, StorageLevel in
pyspark.sql.catalog and pyspark.sql.connect.catalog, and sys in
pyspark.sql.connect.functions.builtin. Importing
pyspark.sql.connect.streaming.readwriter also has to load
pyspark.sql.connect.readwriter, whose namespace the stream reader's doctests
read.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]