luis4a0 commented on PR #12963: URL: https://github.com/apache/gluten/pull/12963#issuecomment-6467991687
@rui-mo, I took the ordinary columnar-input approach rather than retaining the new stage contract. Commit: https://github.com/apache/gluten/commit/424e19ea6d838d7dea655d08294fb596dc400fdf `OneRowRelationExecTransformer` now supplies a normal one-partition RDD that creates one zero-column, one-row batch inside the task. It uses the existing input-iterator relation and empty-schema JNI batch bridge. There is no `supportsNoInputExecution` API or propagation through unary/union/broadcast joins. I also removed the now-unused OneRow JVM serializer and this PR's native ValuesNode scaffolding, restoring those common/native paths to the merged public base. The operator map retains only `OneRowRelationExec` and names the actual native `TableScanNode` iterator connector; the typed VirtualTable row and its support note are gone. The new input-RDD assertion failed before this change and passes afterward. Fresh validation passed: both Spark 4.1/4.2 reactors; 106 Spark 4.1 operator tests (one pre-existing ignored); seven OneRow cases on JDK 21; nine JVM partition/fallback/wire tests; ten native tests; Spark 4.1 EXPLAIN; and all 66 Spark 4.2 plan/runtime tests with golden generation disabled. Rebuilt native libraries were verified inside the tested jars. The PR description now reflects this simpler design. Fresh upstream CI and your review are still needed; these local results do not waive either gate. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
