Spark SQL Catalyst源码分析之Optimizer

浏览数：30 / 时间：2015年06月12日

前几篇文章介绍了Spark SQL的Catalyst的核心运行流程、SqlParser，和Analyzer 以及核心类库TreeNode，本文将详细讲解Spark SQL的Optimizer的优化思想以及Optimizer在Catalyst里的表现方式，并加上自己的实践，对Optimizer有一个直观的认识。

Optimizer的主要职责是将Analyzer给Resolved的Logical Plan根据不同的优化策略Batch，来对语法树进行优化，优化逻辑计划节点(Logical Plan)以及表达式(Expression)，也是转换成物理执行计划的前置。如图：

一、Optimizer

Optimizer这个类是在catalyst里的optimizer包下的唯一一个类，Optimizer的工作方式其实类似Analyzer，因为它们都继承自RuleExecutor[LogicalPlan]，都是执行一系列的Batch操作：

Optimizer里的batches包含了3类优化策略：1、Combine Limits 合并Limits 2、ConstantFolding 常量合并 3、Filter Pushdown 过滤器下推,每个Batch里定义的优化伴随对象都定义在Optimizer里了：

object Optimizer extends RuleExecutor[LogicalPlan] {
  val batches =
    Batch("Combine Limits", FixedPoint(100),
      CombineLimits) ::
    Batch("ConstantFolding", FixedPoint(100),
      NullPropagation,
      ConstantFolding,
      BooleanSimplification,
      SimplifyFilters,
      SimplifyCasts,
      SimplifyCaseConversionExpressions) ::
    Batch("Filter Pushdown", FixedPoint(100),
      CombineFilters,
      PushPredicateThroughProject,
      PushPredicateThroughJoin,
      ColumnPruning) :: Nil
}

另外提一点，Optimizer里不但对Logical Plan进行了优化，而且对Logical Plan中的Expression也进行了优化，所以有必要了解一下Expression相关类，主要是用到了references和outputSet，references主要是Logical Plan节点的输出，而outputSet是Expression的输出：

二、优化策略详解

Optimizer的优化策略不仅有对plan进行transform的，也有对expression进行transform的，究其原理就是遍历树，然后应用优化的Rule，但是注意一点，对Logical Plan transfrom的是先序遍历(pre-order)，而对Expression transfrom的时候是后序遍历(post-order)：

2.1、Batch: Combine Limits

如果出现了2个Limit，则将2个Limit合并为一个，这个要求一个Limit是另一个Limit的grandChild。

 /**
 * Combines two adjacent [[Limit]] operators into one, merging the
 * expressions into one single expression.
 */
object CombineLimits extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case ll @ Limit(le, nl @ Limit(ne, grandChild)) => //ll为当前Limit,le为其expression， nl是ll的grandChild，ne是nl的expression
      Limit(If(LessThan(ne, le), ne, le), grandChild) //expression比较，如果ne比le小则表达式为ne，否则为le
  }
}

给定SQL：val query = sql("select * from (select * from temp_shengli limit 100)a limit 10 ")

scala> query.queryExecution.analyzed
res12: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Limit 10
 Project [key#13,value#14]
  Limit 100
   Project [key#13,value#14]
    MetastoreRelation default, temp_shengli, None

子查询里limit100,外层查询limit10，这里我们当然可以在子查询里不必查那么多，因为外层只需要10个，所以这里会合并Limit10，和Limit100 为 Limit 10。

2.2、Batch: ConstantFolding

这个Batch里包含了Rules：NullPropagation，ConstantFolding，BooleanSimplification，SimplifyFilters，SimplifyCasts，SimplifyCaseConversionExpressions。

2.2.1、Rule：NullPropagation

这里先提一下Literal字面量，它其实是一个能匹配任意基本类型的类。(为下文做铺垫)

object Literal {
  def apply(v: Any): Literal = v match {
    case i: Int => Literal(i, IntegerType)
    case l: Long => Literal(l, LongType)
    case d: Double => Literal(d, DoubleType)
    case f: Float => Literal(f, FloatType)
    case b: Byte => Literal(b, ByteType)
    case s: Short => Literal(s, ShortType)
    case s: String => Literal(s, StringType)
    case b: Boolean => Literal(b, BooleanType)
    case d: BigDecimal => Literal(d, DecimalType)
    case t: Timestamp => Literal(t, TimestampType)
    case a: Array[Byte] => Literal(a, BinaryType)
    case null => Literal(null, NullType)
  }
}

注意Literal是一个LeafExpression，核心方法是eval，给定Row，计算表达式返回值：

case class Literal(value: Any, dataType: DataType) extends LeafExpression {
  override def foldable = true
  def nullable = value == null
  def references = Set.empty
  override def toString = if (value != null) value.toString else "null"
  type EvaluatedType = Any
  override def eval(input: Row):Any = value
}

现在来看一下NullPropagation都做了什么。

NullPropagation是一个能将Expression Expressions替换为等价的Literal值的优化，并且能够避免NULL值在SQL语法树的传播。

/**
 * Replaces [[Expression Expressions]] that can be statically evaluated with
 * equivalent [[Literal]] values. This rule is more specific with
 * Null value propagation from bottom to top of the expression tree.
 */
object NullPropagation extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case q: LogicalPlan => q transformExpressionsUp {
      case e @ Count(Literal(null, _)) => Cast(Literal(0L), e.dataType) //如果count(null)则转化为count(0)
      case e @ Sum(Literal(c, _)) if c == 0 => Cast(Literal(0L), e.dataType)<span style="font-family: Arial;">//如果sum(null)则转化为sum(0)</span>
      case e @ Average(Literal(c, _)) if c == 0 => Literal(0.0, e.dataType)
      case e @ IsNull(c) if !c.nullable => Literal(false, BooleanType)
      case e @ IsNotNull(c) if !c.nullable => Literal(true, BooleanType)
      case e @ GetItem(Literal(null, _), _) => Literal(null, e.dataType)
      case e @ GetItem(_, Literal(null, _)) => Literal(null, e.dataType)
      case e @ GetField(Literal(null, _), _) => Literal(null, e.dataType)
      case e @ Coalesce(children) => {
        val newChildren = children.filter(c => c match {
          case Literal(null, _) => false
          case _ => true
        })
        if (newChildren.length == 0) {
          Literal(null, e.dataType)
        } else if (newChildren.length == 1) {
          newChildren(0)
        } else {
          Coalesce(newChildren)
        }
      }
      case e @ If(Literal(v, _), trueValue, falseValue) => if (v == true) trueValue else falseValue
      case e @ In(Literal(v, _), list) if (list.exists(c => c match {
          case Literal(candidate, _) if candidate == v => true
          case _ => false
        })) => Literal(true, BooleanType)
      // Put exceptional cases above if any
      case e: BinaryArithmetic => e.children match {
        case Literal(null, _) :: right :: Nil => Literal(null, e.dataType)
        case left :: Literal(null, _) :: Nil => Literal(null, e.dataType)
        case _ => e
      }
      case e: BinaryComparison => e.children match {
        case Literal(null, _) :: right :: Nil => Literal(null, e.dataType)
        case left :: Literal(null, _) :: Nil => Literal(null, e.dataType)
        case _ => e
      }
      case e: StringRegexExpression => e.children match {
        case Literal(null, _) :: right :: Nil => Literal(null, e.dataType)
        case left :: Literal(null, _) :: Nil => Literal(null, e.dataType)
        case _ => e
      }
    }
  }
}

给定SQL: val query = sql("select count(null) from temp_shengli where key is not null")

scala> query.queryExecution.analyzed
res6: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Aggregate [], [COUNT(null) AS c0#5L] //这里count的是null
 Filter IS NOT NULL key#7
  MetastoreRelation default, temp_shengli, None

调用NullPropagation

scala> NullPropagation(query.queryExecution.analyzed)
res7: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Aggregate [], [CAST(0, LongType) AS c0#5L]  //优化后为0了
 Filter IS NOT NULL key#7
  MetastoreRelation default, temp_shengli, None

2.2.2、Rule：ConstantFolding

常量合并是属于Expression优化的一种，对于可以直接计算的常量，不用放到物理执行里去生成对象来计算了，直接可以在计划里就计算出来：

 object ConstantFolding extends Rule[LogicalPlan] {
      def apply(plan: LogicalPlan): LogicalPlan = plan transform { //先对plan进行transform
        case q: LogicalPlan => q transformExpressionsDown { //对每个plan的expression进行transform
          // Skip redundant folding of literals.
          case l: Literal => l
          case e if e.foldable => Literal(e.eval(null), e.dataType) //调用eval方法计算结果
        }
      }
    }

给定SQL： val query = sql("select 1+2+3+4 from temp_shengli")

scala> query.queryExecution.analyzed
res23: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [(((1 + 2) + 3) + 4) AS c0#21]  //这里还是常量表达式
 MetastoreRelation default, src, None

优化后：

scala> query.queryExecution.optimizedPlan
res24: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [10 AS c0#21] //优化后，直接合并为10
 MetastoreRelation default, src, None

2.2.3、BooleanSimplification

这个是对布尔表达式的优化，有点像java布尔表达式中的短路判断，不过这个写的倒是很优雅。

看看布尔表达式2边能不能通过只计算1边，而省去计算另一边而提高效率，称为简化布尔表达式。

解释请看我写的注释：

/**
 * Simplifies boolean expressions where the answer can be determined without evaluating both sides.
 * Note that this rule can eliminate expressions that might otherwise have been evaluated and thus
 * is only safe when evaluations of expressions does not result in side effects.
 */
object BooleanSimplification extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case q: LogicalPlan => q transformExpressionsUp {
      case and @ And(left, right) => //如果布尔表达式是AND操作，即exp1 and exp2
        (left, right) match { //（左边表达式，右边表达式）
          case (Literal(true, BooleanType), r) => r // 左边true，返回右边的<span style="font-family: Arial;">bool</span><span style="font-family: Arial;">值</span>
          case (l, Literal(true, BooleanType)) => l //右边true，返回左边的bool值
          case (Literal(false, BooleanType), _) => Literal(false)//左边都false，右边随便，反正是返回false
          case (_, Literal(false, BooleanType)) => Literal(false)//只要有1边是false了，都是false
          case (_, _) => and
        }

      case or @ Or(left, right) =>
        (left, right) match {
          case (Literal(true, BooleanType), _) => Literal(true) //只要左边是true了，不用判断右边都是true
          case (_, Literal(true, BooleanType)) => Literal(true) //只要有一边是true，都返回true
          case (Literal(false, BooleanType), r) => r //希望右边r是true
          case (l, Literal(false, BooleanType)) => l
          case (_, _) => or
        }
    }
  }
}

2.3 Batch: Filter Pushdown

Filter Pushdown下包含了CombineFilters、PushPredicateThroughProject、PushPredicateThroughJoin、ColumnPruning

Ps：感觉Filter Pushdown的名字起的有点不能涵盖全部比如ColumnPruning列裁剪。

2.3.1、Combine Filters

合并两个相邻的Filter,这个和上述Combine Limit差不多。合并2个节点，就可以减少树的深度从而减少重复执行过滤的代价。

/**
 * Combines two adjacent [[Filter]] operators into one, merging the
 * conditions into one conjunctive predicate.
 */
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)
  }
}

给定SQL：val query = sql("select key from (select key from temp_shengli where key >100)a where key > 80 ")

优化前：我们看到一个filter 是另一个filter的grandChild

scala> query.queryExecution.analyzed
res25: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [key#27]
 Filter (key#27 > 80) //filter>80
  Project [key#27]
   Filter (key#27 > 100) //filter>100
    MetastoreRelation default, src, None

优化后：其实filter也可以表达为一个复杂的boolean表达式

scala> query.queryExecution.optimizedPlan
res26: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [key#27]
 Filter ((key#27 > 100) && (key#27 > 80)) //合并为1个
  MetastoreRelation default, src, None

2.3.2 Filter Pushdown

Filter Pushdown，过滤器下推。

原理就是更早的过滤掉不需要的元素来减少开销。

给定SQL：val query = sql("select key from (select * from temp_shengli)a where key>100")

生成的逻辑计划为：

scala> scala> query.queryExecution.analyzed
res29: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [key#31]
 Filter (key#31 > 100) //先select key, value，然后再Filter
  Project [key#31,value#32]
   MetastoreRelation default, src, None

优化后的计划为：

query.queryExecution.optimizedPlan
res30: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [key#31]
 Filter (key#31 > 100) //先filter,然后再select
  MetastoreRelation default, src, None

2.3.3、ColumnPruning

列裁剪用的比较多，就是减少不必要select的某些列。

列裁剪在3种地方可以用：

1、在聚合操作中，可以做列裁剪

2、在join操作中，左右孩子可以做列裁剪

3、合并相邻的Project的列

object ColumnPruning extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    // Eliminate attributes that are not needed to calculate the specified aggregates.
    case a @ Aggregate(_, _, child) if (child.outputSet -- a.references).nonEmpty => ////如果project的outputSet中减去a.references的元素如果不同，那么就将Aggreagte的child替换为a.references
      a.copy(child = Project(a.references.toSeq, child))

    // Eliminate unneeded attributes from either side of a Join.
    case Project(projectList, Join(left, right, joinType, condition)) =>// 消除join的left 和 right孩子的不必要属性，将join的左右子树的列进行裁剪
      // Collect the list of off references required either above or to evaluate the condition.
      val allReferences: Set[Attribute] =
        projectList.flatMap(_.references).toSet ++ condition.map(_.references).getOrElse(Set.empty)

      /** Applies a projection only when the child is producing unnecessary attributes */
      def prunedChild(c: LogicalPlan) =
        if ((c.outputSet -- allReferences.filter(c.outputSet.contains)).nonEmpty) {
          Project(allReferences.filter(c.outputSet.contains).toSeq, c)
        } else {
          c
        }
      Project(projectList, Join(prunedChild(left), prunedChild(right), joinType, condition))

    // Combine adjacent Projects.
    case Project(projectList1, Project(projectList2, child)) => //合并相邻Project的列
      // Create a map of Aliases to their values from the child projection.
      // e.g., 'SELECT ... FROM (SELECT a + b AS c, d ...)' produces Map(c -> Alias(a + b, c)).
      val aliasMap = projectList2.collect {
        case a @ Alias(e, _) => (a.toAttribute: Expression, a)
      }.toMap

      // Substitute any attributes that are produced by the child projection, so that we safely
      // eliminate it.
      // e.g., 'SELECT c + 1 FROM (SELECT a + b AS C ...' produces 'SELECT a + b + 1 ...'
      // TODO: Fix TransformBase to avoid the cast below.
      val substitutedProjection = projectList1.map(_.transform {
        case a if aliasMap.contains(a) => aliasMap(a)
      }).asInstanceOf[Seq[NamedExpression]]

      Project(substitutedProjection, child)

    // Eliminate no-op Projects
    case Project(projectList, child) if child.output == projectList => child
  }
}

分别举三个例子来对应三种情况进行说明：

1、在聚合操作中，可以做列裁剪

给定SQL：val query = sql("SELECT 1+1 as shengli, key from (select key, value from temp_shengli)a group by key")

优化前：

res57: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Aggregate [key#51], [(1 + 1) AS shengli#49,key#51]
 Project [key#51,value#52] //优化前默认select key 和 value两列
  MetastoreRelation default, temp_shengli, None

优化后：

scala> ColumnPruning1(query.queryExecution.analyzed)
MetastoreRelation default, temp_shengli, None
res59: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Aggregate [key#51], [(1 + 1) AS shengli#49,key#51]
 Project [key#51]  //优化后，列裁剪掉了value，只select key
  MetastoreRelation default, temp_shengli, None

2、在join操作中，左右孩子可以做列裁剪

给定SQL：val query = sql("select a.value qween from (select * from temp_shengli) a join (select * from temp_shengli)b on a.key =b.key ")
没有优化之前：

scala> query.queryExecution.analyzed
res51: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [value#42 AS qween#39]
 Join Inner, Some((key#41 = key#43))
  Project [key#41,value#42]  //这里多select了一列，即value
   MetastoreRelation default, temp_shengli, None
  Project [key#43,value#44]  //这里多select了一列，即value
   MetastoreRelation default, temp_shengli, None

优化后：（ColumnPruning2是我自己调试用的）

scala> ColumnPruning2(query.queryExecution.analyzed)
allReferences is -> Set(key#35, key#37)
MetastoreRelation default, temp_shengli, None
MetastoreRelation default, temp_shengli, None
res47: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [key#35 AS qween#33]
 Join Inner, Some((key#35 = key#37))
  Project [key#35]   //经过列裁剪之后，left Child只需要select key这一个列
   MetastoreRelation default, temp_shengli, None
  Project [key#37]   //经过列裁剪之后，right Child只需要select key这一个列
   MetastoreRelation default, temp_shengli, None

3、合并相邻的Project的列，裁剪

给定SQL：val query = sql("SELECT c + 1 FROM (SELECT 1 + 1 as c from temp_shengli ) a ")

优化前：

scala> query.queryExecution.analyzed
res61: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [(c#56 + 1) AS c0#57]
 Project [(1 + 1) AS c#56]
  MetastoreRelation default, temp_shengli, None

优化后：

scala> query.queryExecution.optimizedPlan
res62: org.apache.spark.sql.catalyst.plans.logical.LogicalPlan = 
Project [(2 AS c#56 + 1) AS c0#57] //将子查询里的c 代入到 外层select里的c，直接计算结果
 MetastoreRelation default, temp_shengli, None