WrappedPositionAccessor is not generated but Position2Accessor

[waterlx]

When trying to partition on a field of StringType within a NestedType,

Schema schema = new Schema(
    Types.NestedField.required(1, "int", Types.IntegerType.get()),
    Types.NestedField.required(2, "struct", Types.StructType.of(
        Types.NestedField.required(5, "string", Types.StringType.get())
    ))
);

PartitionSpec partitionSpec = PartitionSpec.builderFor(schema).identity("struct.string").build();

the following exception is thrown when writing to Iceberg:

java.lang.ClassCastException: Cannot cast org.apache.spark.unsafe.types.UTF8String to java.lang.CharSequence
	at java.lang.Class.cast(Class.java:3369)
	at org.apache.iceberg.spark.source.PartitionKey.get(PartitionKey.java:133)
	at org.apache.iceberg.PartitionSpec.get(PartitionSpec.java:150)
	at org.apache.iceberg.PartitionSpec.partitionToPath(PartitionSpec.java:166)
	at org.apache.iceberg.LocationProviders$DefaultLocationProvider.newDataLocation(LocationProviders.java:58)
	at org.apache.iceberg.spark.source.Writer$WriterFactory$OutputFileFactory.newOutputFile(Writer.java:362)
	at org.apache.iceberg.spark.source.Writer$BaseWriter.openCurrent(Writer.java:437)
	at org.apache.iceberg.spark.source.Writer$PartitionedWriter.write(Writer.java:531)
	at org.apache.iceberg.spark.source.Writer$PartitionedWriter.write(Writer.java:498)
	at org.apache.spark.sql.execution.datasources.v2.DataWritingSparkTask$$anonfun$run$3.apply(WriteToDataSourceV2Exec.scala:118)
	at org.apache.spark.sql.execution.datasources.v2.DataWritingSparkTask$$anonfun$run$3.apply(WriteToDataSourceV2Exec.scala:116)
	at org.apache.spark.util.Utils$.tryWithSafeFinallyAndFailureCallbacks(Utils.scala:1394)
	at org.apache.spark.sql.execution.datasources.v2.DataWritingSparkTask$.run(WriteToDataSourceV2Exec.scala:146)
	at org.apache.spark.sql.execution.datasources.v2.WriteToDataSourceV2Exec$$anonfun$doExecute$2.apply(WriteToDataSourceV2Exec.scala:67)
	at org.apache.spark.sql.execution.datasources.v2.WriteToDataSourceV2Exec$$anonfun$doExecute$2.apply(WriteToDataSourceV2Exec.scala:66)

The accessor seems not generated corretly. WrappedPositionAccessor is expected to be used to access that String field but Position2Accessor is generated instead.

Suspect:
In PartitionKey#newAccessor():

} else if (accessor instanceof PositionAccessor) {
return new Position2Accessor(position, size, (PositionAccessor) accessor);
} else if (...)

All sub-classes of PositionAccessor (such as StringAccessor, DecimalAccessor and BytesAccessor) will be wrapped into Position2Accessor, which is not expected. They are supposed to be wrapped into WrappedPositionAccessor.

[waterlx]

To fix that, do we need to be more strict to exclude the sub-class of PositionAccessor when routing to build Position2Accessor?
Accessors and #983 seems have the same issue?

[waterlx]

@rdblue Does it make sense to you?

[rdblue]

Yes, I think you're right. One solution is to not wrap the specialized accessors. We may also be able to keep a transform function so that it is applied by the outer-most accessor. That would keep the two concerns separate.

[waterlx]

I am about to close this issue as PR #1013 is merged into master