Github user lindblombr commented on a diff in the pull request:
https://github.com/apache/spark/pull/21847#discussion_r206740630
--- Diff:
external/avro/src/main/scala/org/apache/spark/sql/avro/AvroSerializer.scala ---
@@ -165,16 +182,118 @@ class AvroSerializer(rootCatalystType: DataType,
rootAvroType: Schema, nullable:
result
}
- private def resolveNullableType(avroType: Schema, nullable: Boolean):
Schema = {
- if (nullable) {
+ // Resolve an Avro union against a supplied DataType, i.e. a LongType
compared against
+ // a ["null", "long"] should return a schema of type Schema.Type.LONG
+ // This function also handles resolving a DataType against unions of 2
or more types, i.e.
+ // an IntType resolves against a ["int", "long", "null"] will correctly
return a schema of
+ // type Schema.Type.LONG
+ private def resolveUnionType(avroType: Schema, catalystType: DataType,
+ nullable: Boolean): Schema = {
+ if (avroType.getType == Type.UNION) {
// avro uses union to represent nullable type.
- val fields = avroType.getTypes.asScala
- assert(fields.length == 2)
- val actualType = fields.filter(_.getType != NULL)
- assert(actualType.length == 1)
+ val fieldTypes = avroType.getTypes.asScala
+
+ // If we're nullable, we need to have at least two types. Cases
with more than two types
+ // are captured in test("read read-write, read-write w/ schema,
read") w/ test.avro input
+ if (nullable && fieldTypes.length < 2) {
+ throw new IncompatibleSchemaException(
+ s"Cannot resolve nullable ${catalystType} against union type
${avroType}")
+ }
+
+ val actualType = catalystType match {
+ case NullType => fieldTypes.filter(_.getType == Type.NULL)
+ case BooleanType => fieldTypes.filter(_.getType == Type.BOOLEAN)
+ case ByteType => fieldTypes.filter(_.getType == Type.INT)
+ case BinaryType =>
+ val at = fieldTypes.filter(x => x.getType == Type.BYTES ||
x.getType == Type.FIXED)
+ if (at.length > 1) {
+ throw new IncompatibleSchemaException(
+ s"Cannot resolve schema of ${catalystType} against union
${avroType.toString}")
+ } else {
+ at
+ }
+ case ShortType | IntegerType => fieldTypes.filter(_.getType ==
Type.INT)
+ case LongType => fieldTypes.filter(_.getType == Type.LONG)
+ case FloatType => fieldTypes.filter(_.getType == Type.FLOAT)
+ case DoubleType => fieldTypes.filter(_.getType == Type.DOUBLE)
+ case d: DecimalType => fieldTypes.filter(_.getType == Type.STRING)
+ case StringType => fieldTypes
+ .filter(x => x.getType == Type.STRING || x.getType == Type.ENUM)
+ case DateType => fieldTypes.filter(x => x.getType == Type.INT ||
x.getType == Type.LONG)
+ case TimestampType => fieldTypes.filter(_.getType == Type.LONG)
+ case ArrayType(et, containsNull) =>
+ // Find array that matches the element type specified
+ fieldTypes.filter(x => x.getType == Type.ARRAY
+ && typeMatchesSchema(et, x.getElementType))
+ case st: StructType => // Find the matching record!
+ val recordTypes = fieldTypes.filter(x => x.getType ==
Type.RECORD)
+ if (recordTypes.length > 1) {
+ throw new IncompatibleSchemaException(
+ "Unions of multiple record types are NOT supported with
user-specified schema")
+ }
+ recordTypes
+ case MapType(kt, vt, valueContainsNull) =>
+ // Find the map that matches the value type. Maps in Avro are
always key type string
+ fieldTypes.filter(x => x.getType == Type.MAP &&
typeMatchesSchema(vt, x.getValueType))
--- End diff --
I'm still not clear, but I'll check the behavior prior to this change to see
1. If kt is NOT StringType, was the case handled previously (some
string-ification)?
2. If so, I'll update this to match existing behavior
3. If not, it seems we'll have to decide if we want to throw the exception,
do the conversion, or something else?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
