huaxingao commented on a change in pull request #32049:
URL: https://github.com/apache/spark/pull/32049#discussion_r638476267
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetUtils.scala
##########
@@ -127,4 +145,292 @@ object ParquetUtils {
file.getName == ParquetFileWriter.PARQUET_COMMON_METADATA_FILE ||
file.getName == ParquetFileWriter.PARQUET_METADATA_FILE
}
+
+ /**
+ * When the partial Aggregates (Max/Min/Count) are pushed down to parquet,
we don't need to
+ * createRowBaseReader to read data from parquet and aggregate at spark
layer. Instead we want
+ * to calculate the partial Aggregates (Max/Min/Count) result using the
statistics information
+ * from parquet footer file, and then construct an InternalRow from these
Aggregate results.
+ *
+ * @return Aggregate results in the format of InternalRow
+ */
+ private[sql] def aggResultToSparkInternalRows(
+ footer: ParquetMetadata,
+ parquetTypes: Seq[PrimitiveType.PrimitiveTypeName],
+ values: Seq[Any],
+ dataSchema: StructType,
+ datetimeRebaseModeInRead: String,
+ int96RebaseModeInRead: String,
+ convertTz: Option[ZoneId]): InternalRow = {
+ val mutableRow = new SpecificInternalRow(dataSchema.fields.map(x =>
x.dataType))
+ val footerFileMetaData = footer.getFileMetaData
+ val datetimeRebaseMode = DataSourceUtils.datetimeRebaseMode(
+ footerFileMetaData.getKeyValueMetaData.get,
+ datetimeRebaseModeInRead)
+ val int96RebaseMode = DataSourceUtils.int96RebaseMode(
+ footerFileMetaData.getKeyValueMetaData.get,
+ int96RebaseModeInRead)
+ parquetTypes.zipWithIndex.map {
+ case (PrimitiveType.PrimitiveTypeName.INT32, i) =>
+ dataSchema.fields(i).dataType match {
+ case ByteType =>
+ mutableRow.setByte(i, values(i).asInstanceOf[Integer].toByte)
+ case ShortType =>
+ mutableRow.setShort(i, values(i).asInstanceOf[Integer].toShort)
+ case IntegerType =>
+ mutableRow.setInt(i, values(i).asInstanceOf[Integer])
+ case DateType =>
+ val dateRebaseFunc = DataSourceUtils.creteDateRebaseFuncInRead(
+ datetimeRebaseMode, "Parquet")
+ mutableRow.update(i,
dateRebaseFunc(values(i).asInstanceOf[Integer]))
+ case d: DecimalType =>
+ val decimal = Decimal(values(i).asInstanceOf[Integer].toLong,
d.precision, d.scale)
+ mutableRow.setDecimal(i, decimal, d.precision)
+ case _ => throw new IllegalArgumentException("Unexpected type for
INT32")
+ }
+ case (PrimitiveType.PrimitiveTypeName.INT64, i) =>
+ dataSchema.fields(i).dataType match {
+ case LongType =>
+ mutableRow.setLong(i, values(i).asInstanceOf[Long])
+ case d: DecimalType =>
+ val decimal = Decimal(values(i).asInstanceOf[Long], d.precision,
d.scale)
+ mutableRow.setDecimal(i, decimal, d.precision)
+ case _ => throw new IllegalArgumentException("Unexpected type for
INT64")
+ }
+ case (PrimitiveType.PrimitiveTypeName.INT96, i) =>
+ dataSchema.fields(i).dataType match {
+ case LongType =>
+ mutableRow.setLong(i, values(i).asInstanceOf[Long])
+ case TimestampType =>
+ val int96RebaseFunc =
DataSourceUtils.creteTimestampRebaseFuncInRead(
+ int96RebaseMode, "Parquet INT96")
+ val julianMicros =
+
ParquetRowConverter.binaryToSQLTimestamp(values(i).asInstanceOf[Binary])
+ val gregorianMicros = int96RebaseFunc(julianMicros)
+ val adjTime =
+ convertTz.map(DateTimeUtils.convertTz(gregorianMicros, _,
ZoneOffset.UTC))
+ .getOrElse(gregorianMicros)
+ mutableRow.setLong(i, adjTime)
+ case _ => throw new IllegalArgumentException("Unexpected type for
INT96")
+ }
+ case (PrimitiveType.PrimitiveTypeName.FLOAT, i) =>
+ mutableRow.setFloat(i, values(i).asInstanceOf[Float])
+ case (PrimitiveType.PrimitiveTypeName.DOUBLE, i) =>
+ mutableRow.setDouble(i, values(i).asInstanceOf[Double])
+ case (PrimitiveType.PrimitiveTypeName.BOOLEAN, i) =>
+ mutableRow.setBoolean(i, values(i).asInstanceOf[Boolean])
+ case (PrimitiveType.PrimitiveTypeName.BINARY, i) =>
+ val bytes = values(i).asInstanceOf[Binary].getBytes
+ dataSchema.fields(i).dataType match {
+ case StringType =>
+ mutableRow.update(i, UTF8String.fromBytes(bytes))
+ case BinaryType =>
+ mutableRow.update(i, bytes)
+ case d: DecimalType =>
+ val decimal =
+ Decimal(new BigDecimal(new BigInteger(bytes), d.scale),
d.precision, d.scale)
+ mutableRow.setDecimal(i, decimal, d.precision)
+ case _ => throw new IllegalArgumentException("Unexpected type for
Binary")
+ }
+ case (PrimitiveType.PrimitiveTypeName.FIXED_LEN_BYTE_ARRAY, i) =>
+ val bytes = values(i).asInstanceOf[Binary].getBytes
+ dataSchema.fields(i).dataType match {
+ case d: DecimalType =>
+ val decimal =
+ Decimal(new BigDecimal(new BigInteger(bytes), d.scale),
d.precision, d.scale)
+ mutableRow.setDecimal(i, decimal, d.precision)
+ case _ => throw new IllegalArgumentException("Unexpected type for
FIXED_LEN_BYTE_ARRAY")
+ }
+ case _ =>
+ throw new IllegalArgumentException("Unexpected parquet type name")
+ }
+ mutableRow
+ }
+
+ /**
+ * When the Aggregates (Max/Min/Count) are pushed down to parquet, in the
case of
+ * PARQUET_VECTORIZED_READER_ENABLED sets to true, we don't need
buildColumnarReader
+ * to read data from parquet and aggregate at spark layer. Instead we want
+ * to calculate the Aggregates (Max/Min/Count) result using the statistics
information
+ * from parquet footer file, and then construct a ColumnarBatch from these
Aggregate results.
+ *
+ * @return Aggregate results in the format of ColumnarBatch
+ */
+ private[sql] def aggResultToSparkColumnarBatch(
+ footer: ParquetMetadata,
+ parquetTypes: Seq[PrimitiveType.PrimitiveTypeName],
+ values: Seq[Any],
+ dataSchema: StructType,
+ offHeap: Boolean,
+ datetimeRebaseModeInRead: String,
+ int96RebaseModeInRead: String,
+ convertTz: Option[ZoneId]): ColumnarBatch = {
+ val capacity = 4 * 1024
+ val footerFileMetaData = footer.getFileMetaData
+ val datetimeRebaseMode = DataSourceUtils.datetimeRebaseMode(
+ footerFileMetaData.getKeyValueMetaData.get,
+ datetimeRebaseModeInRead)
+ val int96RebaseMode = DataSourceUtils.int96RebaseMode(
+ footerFileMetaData.getKeyValueMetaData.get,
+ int96RebaseModeInRead)
+ val columnVectors = if (offHeap) {
+ OffHeapColumnVector.allocateColumns(capacity, dataSchema)
+ } else {
+ OnHeapColumnVector.allocateColumns(capacity, dataSchema)
+ }
+
+ parquetTypes.zipWithIndex.map {
+ case (PrimitiveType.PrimitiveTypeName.INT32, i) =>
+ dataSchema.fields(i).dataType match {
+ case ByteType =>
+ columnVectors(i).appendByte(values(i).asInstanceOf[Integer].toByte)
+ case ShortType =>
+
columnVectors(i).appendShort(values(i).asInstanceOf[Integer].toShort)
+ case IntegerType =>
+ columnVectors(i).appendInt(values(i).asInstanceOf[Integer])
+ case DateType =>
+ val dateRebaseFunc = DataSourceUtils.creteDateRebaseFuncInRead(
+ datetimeRebaseMode, "Parquet")
+
columnVectors(i).appendInt(dateRebaseFunc(values(i).asInstanceOf[Integer]))
+ case _ => throw new IllegalArgumentException("Unexpected type for
INT32")
+ }
+ case (PrimitiveType.PrimitiveTypeName.INT64, i) =>
+ columnVectors(i).appendLong(values(i).asInstanceOf[Long])
+ case (PrimitiveType.PrimitiveTypeName.INT96, i) =>
+ dataSchema.fields(i).dataType match {
+ case LongType =>
+ columnVectors(i).appendLong(values(i).asInstanceOf[Long])
+ case TimestampType =>
+ val int96RebaseFunc =
DataSourceUtils.creteTimestampRebaseFuncInRead(
+ int96RebaseMode, "Parquet INT96")
+ val julianMicros =
+
ParquetRowConverter.binaryToSQLTimestamp(values(i).asInstanceOf[Binary])
+ val gregorianMicros = int96RebaseFunc(julianMicros)
+ val adjTime =
+ convertTz.map(DateTimeUtils.convertTz(gregorianMicros, _,
ZoneOffset.UTC))
+ .getOrElse(gregorianMicros)
+ columnVectors(i).appendLong(adjTime)
+ case _ => throw new IllegalArgumentException("Unexpected type for
INT96")
+ }
+ case (PrimitiveType.PrimitiveTypeName.FLOAT, i) =>
+ columnVectors(i).appendFloat(values(i).asInstanceOf[Float])
+ case (PrimitiveType.PrimitiveTypeName.DOUBLE, i) =>
+ columnVectors(i).appendDouble(values(i).asInstanceOf[Double])
+ case (PrimitiveType.PrimitiveTypeName.BINARY, i) =>
+ val bytes = values(i).asInstanceOf[Binary].getBytes
+ columnVectors(i).putByteArray(0, bytes, 0, bytes.length)
+ case (PrimitiveType.PrimitiveTypeName.FIXED_LEN_BYTE_ARRAY, i) =>
+ val bytes = values(i).asInstanceOf[Binary].getBytes
+ columnVectors(i).putByteArray(0, bytes, 0, bytes.length)
+ case (PrimitiveType.PrimitiveTypeName.BOOLEAN, i) =>
+ columnVectors(i).appendBoolean(values(i).asInstanceOf[Boolean])
+ case _ =>
+ throw new IllegalArgumentException("Unexpected parquet type name")
+ }
+ new ColumnarBatch(columnVectors.asInstanceOf[Array[ColumnVector]], 1)
+ }
+
+ /**
+ * Calculate the pushed down Aggregates (Max/Min/Count) result using the
statistics
+ * information from parquet footer file.
+ *
+ * @return A tuple of `Array[PrimitiveType.PrimitiveTypeName]` and
Array[Any].
+ * The first element is the PrimitiveTypeName of the Aggregate
column,
+ * and the second element is the aggregated value.
+ */
+ private[sql] def getPushedDownAggResult(
+ footer: ParquetMetadata,
+ dataSchema: StructType,
+ aggregation: Aggregation)
+ : (Array[PrimitiveType.PrimitiveTypeName], Array[Any]) = {
+ val footerFileMetaData = footer.getFileMetaData
+ val fields = footerFileMetaData.getSchema.getFields
+ val blocks = footer.getBlocks()
+ val typesBuilder = ArrayBuilder.make[PrimitiveType.PrimitiveTypeName]
+ val valuesBuilder = ArrayBuilder.make[Any]
+
+ for (i <- 0 until aggregation.aggregateExpressions.size) {
+ var value: Any = None
+ var rowCount = 0L
+ var isCount = false
+ var index = 0
+ blocks.forEach { block =>
+ val blockMetaData = block.getColumns()
+ aggregation.aggregateExpressions(i) match {
+ case Seq(Max(col, _)) =>
+ index = dataSchema.fieldNames.toList.indexOf(col)
+ val currentMax = getCurrentBlockMaxOrMin(footer, blockMetaData,
index, true)
+ if (currentMax != None &&
+ (value == None ||
currentMax.asInstanceOf[Comparable[Any]].compareTo(value) > 0)) {
+ value = currentMax
+ }
+
+ case Seq(Min(col, _)) =>
+ index = dataSchema.fieldNames.toList.indexOf(col)
+ val currentMin = getCurrentBlockMaxOrMin(footer, blockMetaData,
index, false)
+ if (currentMin != None &&
+ (value == None ||
currentMin.asInstanceOf[Comparable[Any]].compareTo(value) < 0)) {
+ value = currentMin
+ }
+
+ case Seq(Count(col, _, _)) =>
+ index = dataSchema.fieldNames.toList.indexOf(col)
+ rowCount += block.getRowCount
+ if (!col.equals("1")) { // "1" is for count(*)
+ rowCount -= getNumNulls(footer, blockMetaData, index)
+ }
+ isCount = true
+
+ case _ =>
+ }
+ }
+ if (isCount) {
+ valuesBuilder += rowCount
+ typesBuilder += PrimitiveType.PrimitiveTypeName.INT96
+ } else {
+ valuesBuilder += value
+ typesBuilder += fields.get(index).asPrimitiveType.getPrimitiveTypeName
+ }
+ }
+ (typesBuilder.result(), valuesBuilder.result())
+ }
+
+ /**
+ * get the Max or Min value for ith column in the current block
+ *
+ * @return the Max or Min value
+ */
+ private def getCurrentBlockMaxOrMin(
+ footer: ParquetMetadata,
+ columnChunkMetaData: util.List[ColumnChunkMetaData],
+ i: Int,
+ isMax: Boolean): Any = {
+ val parquetType = footer.getFileMetaData.getSchema.getType(i)
+ if (!parquetType.isPrimitive) {
+ throw new IllegalArgumentException("Unsupported type : " +
parquetType.toString)
+ }
+ val statistics = columnChunkMetaData.get(i).getStatistics()
+ if (!statistics.hasNonNullValue) {
+ throw new UnsupportedOperationException("No min/max found for parquet
file, Set SQLConf" +
+ " PARQUET_AGGREGATE_PUSHDOWN_ENABLED to false and execute again")
Review comment:
This part bothers me too. I don't know why in some cases parquet doesn't
return statistics info. We won't be able to find out until we read footer in
`FilePartitionReaderFactory.createReader`. It's already passed the
`SupportsPushDownAggregates .pushAggregation`. I can't figure out a good way to
do this.
--
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]
