msamirkhan commented on a change in pull request #29353:
URL: https://github.com/apache/spark/pull/29353#discussion_r465870192
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/orc/OrcDeserializer.scala
##########
@@ -73,136 +74,180 @@ class OrcDeserializer(
* Creates a writer to write ORC values to Catalyst data structure at the
given ordinal.
*/
private def newWriter(
- dataType: DataType, updater: CatalystDataUpdater): (Int,
WritableComparable[_]) => Unit =
+ dataType: DataType): (CatalystDataUpdater, Int, WritableComparable[_])
=> Unit =
dataType match {
- case NullType => (ordinal, _) =>
+ case NullType => (updater, ordinal, _) =>
updater.setNullAt(ordinal)
- case BooleanType => (ordinal, value) =>
+ case BooleanType => (updater, ordinal, value) =>
updater.setBoolean(ordinal, value.asInstanceOf[BooleanWritable].get)
- case ByteType => (ordinal, value) =>
+ case ByteType => (updater, ordinal, value) =>
updater.setByte(ordinal, value.asInstanceOf[ByteWritable].get)
- case ShortType => (ordinal, value) =>
+ case ShortType => (updater, ordinal, value) =>
updater.setShort(ordinal, value.asInstanceOf[ShortWritable].get)
- case IntegerType => (ordinal, value) =>
+ case IntegerType => (updater, ordinal, value) =>
updater.setInt(ordinal, value.asInstanceOf[IntWritable].get)
- case LongType => (ordinal, value) =>
+ case LongType => (updater, ordinal, value) =>
updater.setLong(ordinal, value.asInstanceOf[LongWritable].get)
- case FloatType => (ordinal, value) =>
+ case FloatType => (updater, ordinal, value) =>
updater.setFloat(ordinal, value.asInstanceOf[FloatWritable].get)
- case DoubleType => (ordinal, value) =>
+ case DoubleType => (updater, ordinal, value) =>
updater.setDouble(ordinal, value.asInstanceOf[DoubleWritable].get)
- case StringType => (ordinal, value) =>
+ case StringType => (updater, ordinal, value) =>
updater.set(ordinal,
UTF8String.fromBytes(value.asInstanceOf[Text].copyBytes))
- case BinaryType => (ordinal, value) =>
+ case BinaryType => (updater, ordinal, value) =>
val binary = value.asInstanceOf[BytesWritable]
val bytes = new Array[Byte](binary.getLength)
System.arraycopy(binary.getBytes, 0, bytes, 0, binary.getLength)
updater.set(ordinal, bytes)
- case DateType => (ordinal, value) =>
+ case DateType => (updater, ordinal, value) =>
updater.setInt(ordinal, OrcShimUtils.getGregorianDays(value))
- case TimestampType => (ordinal, value) =>
+ case TimestampType => (updater, ordinal, value) =>
updater.setLong(ordinal,
DateTimeUtils.fromJavaTimestamp(value.asInstanceOf[OrcTimestamp]))
- case DecimalType.Fixed(precision, scale) => (ordinal, value) =>
+ case DecimalType.Fixed(precision, scale) => (updater, ordinal, value) =>
val v = OrcShimUtils.getDecimal(value)
v.changePrecision(precision, scale)
updater.set(ordinal, v)
- case st: StructType => (ordinal, value) =>
- val result = new SpecificInternalRow(st)
- val fieldUpdater = new RowUpdater(result)
+ case st: StructType =>
+ val createRow: () => InternalRow = getRowCreator(st)
val fieldConverters = st.map(_.dataType).map { dt =>
- newWriter(dt, fieldUpdater)
+ newWriter(dt)
}.toArray
- val orcStruct = value.asInstanceOf[OrcStruct]
-
- var i = 0
- while (i < st.length) {
- val value = orcStruct.getFieldValue(i)
- if (value == null) {
- result.setNullAt(i)
- } else {
- fieldConverters(i)(i, value)
+ (updater, ordinal, value) => {
+ val result = createRow.apply()
+ val fieldUpdater = new RowUpdater(result)
+ val orcStruct = value.asInstanceOf[OrcStruct]
+
+ var i = 0
+ while (i < st.length) {
+ val value = orcStruct.getFieldValue(i)
+ if (value == null) {
+ result.setNullAt(i)
+ } else {
+ fieldConverters(i)(fieldUpdater, i, value)
+ }
+ i += 1
}
- i += 1
+
+ updater.set(ordinal, result)
}
- updater.set(ordinal, result)
-
- case ArrayType(elementType, _) => (ordinal, value) =>
- val orcArray = value.asInstanceOf[OrcList[WritableComparable[_]]]
- val length = orcArray.size()
- val result = createArrayData(elementType, length)
- val elementUpdater = new ArrayDataUpdater(result)
- val elementConverter = newWriter(elementType, elementUpdater)
-
- var i = 0
- while (i < length) {
- val value = orcArray.get(i)
- if (value == null) {
- result.setNullAt(i)
- } else {
- elementConverter(i, value)
+ case ArrayType(elementType, _) =>
+ val elementConverter = newWriter(elementType)
+ val createArray = getArrayDataCreator(elementType)
+ (updater, ordinal, value) => {
+ val orcArray = value.asInstanceOf[OrcList[WritableComparable[_]]]
+ val length = orcArray.size()
+ val result = createArray(length)
+ val elementUpdater = new ArrayDataUpdater(result)
+
+ var i = 0
+ while (i < length) {
+ val value = orcArray.get(i)
+ if (value == null) {
+ result.setNullAt(i)
+ } else {
+ elementConverter(elementUpdater, i, value)
+ }
+ i += 1
}
- i += 1
+
+ updater.set(ordinal, result)
}
- updater.set(ordinal, result)
-
- case MapType(keyType, valueType, _) => (ordinal, value) =>
- val orcMap = value.asInstanceOf[OrcMap[WritableComparable[_],
WritableComparable[_]]]
- val length = orcMap.size()
- val keyArray = createArrayData(keyType, length)
- val keyUpdater = new ArrayDataUpdater(keyArray)
- val keyConverter = newWriter(keyType, keyUpdater)
- val valueArray = createArrayData(valueType, length)
- val valueUpdater = new ArrayDataUpdater(valueArray)
- val valueConverter = newWriter(valueType, valueUpdater)
-
- var i = 0
- val it = orcMap.entrySet().iterator()
- while (it.hasNext) {
- val entry = it.next()
- keyConverter(i, entry.getKey)
- val value = entry.getValue
- if (value == null) {
- valueArray.setNullAt(i)
- } else {
- valueConverter(i, value)
+ case MapType(keyType, valueType, _) =>
+ val keyConverter = newWriter(keyType)
+ val createKeyArray = getArrayDataCreator(keyType)
+ val valueConverter = newWriter(valueType)
+ val createValueArray = getArrayDataCreator(valueType)
+ (updater, ordinal, value) => {
+ val orcMap = value.asInstanceOf[OrcMap[WritableComparable[_],
WritableComparable[_]]]
+ val length = orcMap.size()
+ val keyArray = createKeyArray(length)
+ val keyUpdater = new ArrayDataUpdater(keyArray)
+ val valueArray = createValueArray(length)
+ val valueUpdater = new ArrayDataUpdater(valueArray)
+
+ var i = 0
+ val it = orcMap.entrySet().iterator()
+ while (it.hasNext) {
+ val entry = it.next()
+ keyConverter(keyUpdater, i, entry.getKey)
+ val value = entry.getValue
+ if (value == null) {
+ valueArray.setNullAt(i)
+ } else {
+ valueConverter(valueUpdater, i, value)
+ }
+ i += 1
}
- i += 1
- }
- // The ORC map will never have null or duplicated map keys, it's safe
to create a
- // ArrayBasedMapData directly here.
- updater.set(ordinal, new ArrayBasedMapData(keyArray, valueArray))
+ // The ORC map will never have null or duplicated map keys, it's
safe to create a
+ // ArrayBasedMapData directly here.
+ updater.set(ordinal, new ArrayBasedMapData(keyArray, valueArray))
+ }
- case udt: UserDefinedType[_] => newWriter(udt.sqlType, updater)
+ case udt: UserDefinedType[_] =>
+ val converter = newWriter(udt.sqlType)
+ (updater, ordinal, value) => {
+ converter(updater, ordinal, value)
+ }
case _ =>
throw new UnsupportedOperationException(s"$dataType is not supported
yet.")
}
- private def createArrayData(elementType: DataType, length: Int): ArrayData =
elementType match {
- case BooleanType => UnsafeArrayData.fromPrimitiveArray(new
Array[Boolean](length))
- case ByteType => UnsafeArrayData.fromPrimitiveArray(new
Array[Byte](length))
- case ShortType => UnsafeArrayData.fromPrimitiveArray(new
Array[Short](length))
- case IntegerType => UnsafeArrayData.fromPrimitiveArray(new
Array[Int](length))
- case LongType => UnsafeArrayData.fromPrimitiveArray(new
Array[Long](length))
- case FloatType => UnsafeArrayData.fromPrimitiveArray(new
Array[Float](length))
- case DoubleType => UnsafeArrayData.fromPrimitiveArray(new
Array[Double](length))
- case _ => new GenericArrayData(new Array[Any](length))
+ private def getArrayDataCreator(elementType: DataType): Int => ArrayData =
elementType match {
+ case BooleanType => length => UnsafeArrayData.createFreshArray(length, 1)
+ case ByteType => length => UnsafeArrayData.createFreshArray(length, 1)
+ case ShortType => length => UnsafeArrayData.createFreshArray(length, 2)
+ case IntegerType => length => UnsafeArrayData.createFreshArray(length, 4)
+ case LongType => length => UnsafeArrayData.createFreshArray(length, 8)
+ case FloatType => length => UnsafeArrayData.createFreshArray(length, 4)
+ case DoubleType => length => UnsafeArrayData.createFreshArray(length, 8)
+ case _ => length => new GenericArrayData(new Array[Any](length))
+ }
+
+ private def getRowCreator(st: StructType): () => InternalRow = {
+ val constructorsArray = new Array[() => MutableValue](st.fields.length)
+ var i = 0
+ while (i < st.fields.length) {
+ st.fields(i).dataType match {
+ case BooleanType => constructorsArray(i) = () => new MutableBoolean
+ case ByteType => constructorsArray(i) = () => new MutableByte
+ case ShortType => constructorsArray(i) = () => new MutableShort
+ // We use INT for DATE internally
+ case IntegerType | DateType => constructorsArray(i) = () => new
MutableInt
+ // We use Long for Timestamp internally
+ case LongType | TimestampType => constructorsArray(i) = () => new
MutableLong
+ case FloatType => constructorsArray(i) = () => new MutableFloat
+ case DoubleType => constructorsArray(i) = () => new MutableDouble
+ case _ => constructorsArray(i) = () => new MutableAny
+ }
+ i += 1
+ }
+
+ () => {
+ val array = new Array[MutableValue](constructorsArray.length)
+ var i = 0
+ while (i < constructorsArray.length) {
+ array(i) = constructorsArray(i).apply()
+ i += 1
+ }
+ new SpecificInternalRow(array)
+ }
Review comment:
While working with avro (https://github.com/apache/spark/pull/29354) the
profiler showed some time being spent in SpecificInternalRow constructor, and
we saw improvements when moving to this model (please see the pdf in the avro
PR) where based on the schema we can fill in a constructors array and for each
data point, call these constructors one by one.
In retrospect, I did not check if the cost overhead is because of two Scala
map calls in SpecificInternalRow constructors or because of a lot of case/match
calls. I will investigate a bit.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
