Details
-
Sub-task
-
Status: Open
-
Major
-
Resolution: Unresolved
-
3.4.0
-
None
-
None
Description
SPARK-35800 introduced a testable instance of GroupState in Scala/Java API. The rationalization was that default implementation of GroupState couldn't be instantiated.
The GroupState being introduced in PySpark is something end users can create an instance directly, but looks like the constructor is only accounting the case it comes from serialized form from executor.
It would be better to enable GroupState instance to be created by end user for testing purpose, with isolation to the constructor for internal usage.