通过Z-Order技术加速Hudi大规模数据集分析方案

x value：01100001
y value：11010110

z-value： 1011011000101001

/** Generates z-value*/
val newRDD = df.rdd.map { row =>
  val values = zFields.map { case (index, field) =>
    field.dataType match {
      case LongType =>
        ZOrderingUtil.longTo8Byte(row.getLong(index))
      case DoubleType =>
        ZOrderingUtil.doubleTo8Byte(row.getDouble(index))
      case IntegerType =>
        ZOrderingUtil.intTo8Byte(row.getInt(index))
      case FloatType =>
        ZOrderingUtil.doubleTo8Byte(row.getFloat(index).toDouble)
      case StringType =>
        ZOrderingUtil.utf8To8Byte(row.getString(index))
      case DateType =>
        ZOrderingUtil.longTo8Byte(row.getDate(index).getTime)
      case TimestampType =>
        ZOrderingUtil.longTo8Byte(row.getTimestamp(index).getTime)
      case ByteType =>
        ZOrderingUtil.byteTo8Byte(row.getByte(index))
      case ShortType =>
        ZOrderingUtil.intTo8Byte(row.getShort(index).toInt)
      case d: DecimalType =>
        ZOrderingUtil.longTo8Byte(row.getDecimal(index).longValue())
      case _ =>
        null
    }
  }.filter(v => v != null).toArray
  val zValues = ZOrderingUtil.interleaveMulti8Byte(values)
  Row.fromSeq(row.toSeq ++ Seq(zValues))
}.sortBy(x => ZorderingBinarySort(x.getAs[Array[Byte]](fieldNum)))

/** Generates z-value */
val indexRdd = internalRdd.mapPartitionsInternal { iter =>
  val bounds = boundBroadCast.value
  val origin_Projections = sortingExpressions.map { se =>
    UnsafeProjection.create(Seq(se), outputAttributes)
  }
  iter.map { unsafeRow =>
    val interleaveValues = origin_Projections.zip(origin_lazyGeneratedOrderings).zipWithIndex.map { case ((rowProject, lazyOrdering), index) =>
      val row = rowProject(unsafeRow)
      val decisionBound = new DecisionBound(sampleRdd, lazyOrdering)
      if (row.isNullAt(0)) {
        bounds(index).length + 1
      } else {
        decisionBound.getBound(row, bounds(index).asInstanceOf[Array[InternalRow]])
      }
    }.toArray.map(ZOrderingUtil.toBytes(_))
    val zValues = ZOrderingUtil.interleaveMulti4Byte(interleaveValues)
    val mutablePair = new MutablePair[InternalRow, Array[Byte]]()
    mutablePair.update(unsafeRow, zValues)
  }
}.sortBy(x => ZorderingBinarySort(x._2), numPartitions = fileNum).map(_._1)

/** collect statistic info*/
val sc = df.sparkSession.sparkContext
val serializableConfiguration = new SerializableConfiguration(conf)
val numParallelism = inputFiles.size/3
val previousJobDescription = sc.getLocalProperty(SparkContext.SPARK_JOB_DESCRIPTION)
try {
  val description = s"Listing parquet column statistics"
  sc.setJobDescription(description)
  sc.parallelize(inputFiles, numParallelism).mapPartitions { paths =>
    val hadoopConf = serializableConfiguration.value
    paths.map(new Path(_)).flatMap { filePath =>
      val blocks = ParquetFileReader.readFooter(hadoopConf, filePath).getBlocks().asScala
      blocks.flatMap(b => b.getColumns().asScala.
        map(col => (col.getPath().toDotString(),
          FileStats(col.getStatistics().minAsString(), col.getStatistics().maxAsString(), col.getStatistics.getNumNulls.toInt))))
        .groupBy(x => x._1).mapValues(v => v.map(vv => vv._2)).
        mapValues(value => FileStats(value.map(_.minVal).min, value.map(_.maxVal).max, value.map(_.num_nulls).max)).toSeq.
        map(x => ColumnFileStats(filePath.getName(), x._1, x._2.minVal, x._2.maxVal, x._2.num_nulls))
    }.filter(p => cols.contains(p.colName))
  }.collect()
} finally {
  sc.setJobDescription(previousJobDescription)
}

/** collect statistic info*/
val inputFiles = df.inputFiles
val conf = df.sparkSession.sparkContext.hadoopConfiguration
val values = cols.flatMap(c => Seq( min(col(c)).as(c + "_minValue"), max(col(c)).as(c + "_maxValue"), count(c).as(c + "_noNullCount")))
val valueCounts = count("*").as("totalNum")
val projectValues = Seq(col("file")) ++ cols.flatMap(c =>
  Seq(col(c + "_minValue"), col(c + "_maxValue"), expr(s"totalNum - ${c + "_noNullCount"}").as(c + "_num_nulls")))
val result = df.select(input_file_name() as "file", col("*"))
  .groupBy($"file")
  .agg(valueCounts,  values: _*).select(projectValues:_*)
result

/** convert filter */
def createZindexFilter(condition: Expression): Expression = {
  val minValue = (colName: Seq[String]) =>
    col(UnresolvedAttribute(colName) + "_minValue").expr
  val maxValue = (colName: Seq[String]) =>
    col(UnresolvedAttribute(colName) + "_maxValue").expr
  val num_nulls = (colName: Seq[String]) =>
    col(UnresolvedAttribute(colName) + "_num_nulls").expr
  condition match {
    case EqualTo(attribute: AttributeReference, value: Literal) =>
      val colName = HudiMergeIntoUtils.getTargetColNameParts(attribute)
      And(LessThanOrEqual(minValue(colName), value), GreaterThanOrEqual(maxValue(colName), value))
    case EqualTo(value: Literal, attribute: AttributeReference) =>
      val colName = HudiMergeIntoUtils.getTargetColNameParts(attribute)
      And(LessThanOrEqual(minValue(colName), value), GreaterThanOrEqual(maxValue(colName), value))
    case equalNullSafe @ EqualNullSafe(_: AttributeReference, _ @ Literal(null, _)) =>
      val colName = HudiMergeIntoUtils.getTargetColNameParts(equalNullSafe.left)
      EqualTo(num_nulls(colName), equalNullSafe.right)
.......