Catalyst 优化逻辑执行计划规则

Optimizer

本文分析Catalyst Optimize部分实现的对逻辑执行计划(LogicalPlan)的处理规则。

Optimizer处理的是LogicalPlan对象。

Optimizer的batches如下：

object Optimizer extends RuleExecutor[LogicalPlan] {
  val batches =
    Batch("ConstantFolding", Once,
      ConstantFolding, // 可静态分析的常量表达式
      BooleanSimplification, // 布尔表达式提前短路
      SimplifyFilters, // 简化过滤操作(false, true, null)
      SimplifyCasts) :: // 简化转换(对象所属类已经是Cast目标类)
    Batch("Filter Pushdown", Once,
      CombineFilters, // 相邻(上下级)Filter操作合并
      PushPredicateThroughProject, // 映射操作中的Filter谓词下推
      PushPredicateThroughInnerJoin) :: Nil // inner join操作谓词下推
}

这是4.1号最新的Catalyst  Optimizer的代码。

ConstantFolding 

把可以静态分析出结果的表达式替换成Literal表达式。

object ConstantFolding extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case q: LogicalPlan => q transformExpressionsDown {
      // Skip redundant folding of literals.
      case l: Literal => l
      case e if e.foldable => Literal(e.apply(null), e.dataType)
    }
  }
}

Literal能处理的类型包括Int, Long, Double, Float, Byte,Short, String, Boolean, null。这些类型分别对应的是Catalyst框架的DataType，包括IntegerType, LongType, DoubleType,FloatType, ByteType, ShortType, StringType, BooleanType, NullType。

普通的Literal是不可变的，还有一个可变的MutalLiteral类，有update方法可以改变里面的value。

BooleanSimplification 

提前短路可以短路的布尔表达式

object BooleanSimplification extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case q: LogicalPlan => q transformExpressionsUp {
      case and @ And(left, right) =>
        (left, right) match {
          case (Literal(true, BooleanType), r) => r
          case (l, Literal(true, BooleanType)) => l
          case (Literal(false, BooleanType), _) => Literal(false)
          case (_, Literal(false, BooleanType)) => Literal(false)
          case (_, _) => and
        }

      case or @ Or(left, right) =>
        (left, right) match {
          case (Literal(true, BooleanType), _) => Literal(true)
          case (_, Literal(true, BooleanType)) => Literal(true)
          case (Literal(false, BooleanType), r) => r
          case (l, Literal(false, BooleanType)) => l
          case (_, _) => or
        }
    }
  }
}

SimplifyFilters 

提前处理可以被判断的过滤操作

object SimplifyFilters extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case Filter(Literal(true, BooleanType), child) =>
      child
    case Filter(Literal(null, _), child) =>
      LocalRelation(child.output)
    case Filter(Literal(false, BooleanType), child) =>
      LocalRelation(child.output)
  }
}

SimplifyCasts 

把已经是目标类的Cast表达式替换掉

object SimplifyCasts extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transformAllExpressions {
    case Cast(e, dataType) if e.dataType == dataType => e
  }
}

CombineFilters 

相邻都是过滤操作的话，把两个过滤操作合起来。相邻指的是上下两级。

object CombineFilters extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case ff @ Filter(fc, nf @ Filter(nc, grandChild)) => Filter(And(nc, fc), grandChild)
  }
}

PushPredicateThroughProject 

把Project操作中的过滤操作下推。这一步里顺带做了别名转换的操作（认为开销不大的前提下）。

object PushPredicateThroughProject extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case filter @ Filter(condition, project @ Project(fields, grandChild)) =>
      val sourceAliases = fields.collect { case a @ Alias(c, _) =>
        (a.toAttribute: Attribute) -> c
      }.toMap // 把fields中的别名属性都取出来
      project.copy(child = filter.copy( // 生成新的Filter操作
        replaceAlias(condition, sourceAliases), // condition中有别名的替换掉
        grandChild))
  }

  def replaceAlias(condition: Expression, sourceAliases: Map[Attribute, Expression]): Expression = {
    condition transform {
      case a: AttributeReference => sourceAliases.getOrElse(a, a)
    }
  }
}

PushPredicateThroughInnerJoin 

先找到Filter操作，若Filter操作里面是一次inner join，那么先把Filter条件和inner join条件先全部取出来，

然后把只涉及到左侧或右侧的过滤操作下推到join外部，把剩下来不能下推的条件放到join操作的condition里。

object PushPredicateThroughInnerJoin extends Rule[LogicalPlan] with PredicateHelper {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case f @ Filter(filterCondition, Join(left, right, Inner, joinCondition)) =>
      // 这一步是把过滤条件和join条件里的condition都提取出来
      val allConditions = splitConjunctivePredicates(filterCondition) ++
        joinCondition.map(splitConjunctivePredicates).getOrElse(Nil)
      
      // 把参考属性都属于右侧输出属性的condition挑选到rightCondition里
      val (rightConditions, leftOrJoinConditions) =
        allConditions.partition(_.references subsetOf right.outputSet)
      // 同理，把剩余condition里面，参考属性都属于左侧输出属性的condition挑选到
      // leftCondition里，剩余的就属于joinCondition
      val (leftConditions, joinConditions) =
        leftOrJoinConditions.partition(_.references subsetOf left.outputSet)

      // 生成新的left和right：先把condition里的操作用AND折叠起来，然后将该折叠后的表达式和原始的left/right logical plan合起来生成新的Filter操作，即新的Fil      // ter logical plan
      // 这样就做到了把过滤条件中的谓词下推到了left/right里，即本次inner join的“外部”
      val newLeft = leftConditions.reduceLeftOption(And).map(Filter(_, left)).getOrElse(left)
      val newRight = rightConditions.reduceLeftOption(And).map(Filter(_, right)).getOrElse(right)
      Join(newLeft, newRight, Inner, joinConditions.reduceLeftOption(And))
  }
}

以下帮助理解上面这段代码。

Join操作(LogicalPlan的Binary)

case class Join(
  left: LogicalPlan,
  right: LogicalPlan,
  joinType: JoinType,
  condition: Option[Expression]) extends BinaryNode {

  def references = condition.map(_.references).getOrElse(Set.empty)
  def output = left.output ++ right.output
}

Filter操作(LogicalPlan的Unary)

case class Filter(condition: Expression, child: LogicalPlan) extends UnaryNode {
  def output = child.output
  def references = condition.references
}

reduceLeftOption逻辑是这样的：

def reduceLeftOption[B >: A](op: (B, A) => B): Option[B] =
    if (isEmpty) None else Some(reduceLeft(op))

reduceLeft(op)的结果是op( op( ... op(x_1, x_2) ...,x_{n-1}), x_n)

谓词助手这个trait，负责把And操作里的condition分离开，返回表达式Seq

trait PredicateHelper {
  def splitConjunctivePredicates(condition: Expression): Seq[Expression] = condition match {
    case And(cond1, cond2) => splitConjunctivePredicates(cond1) ++ splitConjunctivePredicates(cond2)
    case other => other :: Nil
  }
}

Example

case class Person(name:String, age: Int)

case classNum(v1: Int, v2: Int)

case one

SELECT  people.age, num.v1,  num.v2

FROM

    people

    JOIN  num

    ON   people.age > 20  and  num.v1> 0

WHERE  num.v2< 50

== QueryPlan ==

Project [age#1:1,v1#2:2,v2#3:3]

CartesianProduct

      Filter(age#1:1 > 20)

          ExistingRdd[name#0,age#1], MappedRDD[4] at map at basicOperators.scala:124

      Filter((v2#3:1 < 50) && (v1#2:0 > 0))

          ExistingRdd [v1#2,v2#3],MappedRDD[10] at map at basicOperators.scala:124

 

分析：where条件 num.v2 < 50 下推到Join里

case two

SELECT people.age,  1+2

FROM

    people

    JOIN  num

    ON   people.name<>’abc’  and  num.v1> 0

WHERE num.v2 < 50

 

== QueryPlan ==

Project [age#1:1,3 AS c1#14]

    CartesianProduct

        Filter NOT(name#0:0 = abc)

            ExistingRdd[name#0,age#1], MappedRDD[4] at map at basicOperators.scala:124

        Filter((v2#3:1 < 50) && (v1#2:0 > 0))

            ExistingRdd[v1#2,v2#3], MappedRDD[10] at map at basicOperators.scala:124

 

分析：1+2 被提前常量折叠，并被取了一个别名

全文完 :)