isthmus: visit(NamedWrite) ignores output_mode, so a write declaring MODIFIED_RECORDS converts to a ROWCOUNT TableModify #1213

Activity

1. nielspardon commented on Oct 5, 2026

   

   nielspardon

   on Oct 5, 2026

   MemberAuthor

   More actions

   Related, on the CTAS path: handleCreateTableAs does check the output mode, but it accepts only OUTPUT_MODE_NO_OUTPUT, and this repo's Spark module writes every CTAS with OUTPUT_MODE_UNSPECIFIED. So a Spark CREATE TABLE t AS SELECT ... plan is refused by isthmus with Can only handle CTAS NamedWrite with output mode NO_OUTPUT, given UNSPECIFIED, even though its create mode (ERROR_IF_EXISTS / IGNORE_IF_EXISTS / REPLACE_IF_EXISTS) is now supported after #1266. The spec doesn't say what an unspecified output mode means for CTAS, so whichever fix lands here should decide whether isthmus reads it as NO_OUTPUT or whether Spark should emit NO_OUTPUT.

   substrait-java/isthmus/src/main/java/io/substrait/isthmus/SubstraitRelNodeConverter.java

   Lines 1210 to 1215 in d05cd6b

   	if (namedWrite.getOutputMode() != AbstractWriteRel.OutputMode.NO_OUTPUT) {
   	throw new UnsupportedOperationException(
   	String.format(
   	"Can only handle CTAS NamedWrite with output mode %s, given %s",
   	AbstractWriteRel.OutputMode.NO_OUTPUT, namedWrite.getOutputMode()));
   	}

   substrait-java/spark/src/main/scala/io/substrait/spark/logical/ToSubstraitRel.scala

   Lines 664 to 676 in d05cd6b

   	private def convertCTAS(
   	table: CatalogTable,
   	mode: SaveMode,
   	query: LogicalPlan,
   	outputColumnNames: Seq[String]): relation.NamedWrite =
   	relation.NamedWrite
   	.builder()
   	.input(visit(query))
   	.operation(WriteOp.CTAS)
   	.outputMode(OutputMode.UNSPECIFIED)
   	.createMode(createMode(mode))
   	.names(table.identifier.unquotedString.split("\\.").toList.asJava)
   	.tableSchema(outputSchema(query.output, outputColumnNames))