Spark SQL Catalyst源代码分析之Analyzer

/** Spark SQL源代码分析系列文章*/

前面几篇文章解说了Spark SQL的核心运行流程和Spark SQL的Catalyst框架的Sql Parser是如何接受用户输入sql，经过解析生成Unresolved Logical Plan的。

我们记得Spark SQL的运行流程中还有一个核心的组件式Analyzer，本文将会介绍Analyzer在Spark SQL里起到了什么作用。

Analyzer位于Catalyst的analysis package下。主要职责是将Sql Parser 未能Resolved的Logical Plan 给Resolved掉。

watermark/2/text/aHR0cDovL2Jsb2cuY3Nkbi5uZXQvb29wc29vbQ==/font/5a6L5L2T/fontsize/400/fill/I0JBQkFCMA==/dissolve/70/gravity/Center" alt="" />

一、Analyzer构造

Analyzer会使用Catalog和FunctionRegistry将UnresolvedAttribute和UnresolvedRelation转换为catalyst里全类型的对象。

Analyzer里面有fixedPoint对象，一个Seq[Batch].

class Analyzer(catalog: Catalog, registry: FunctionRegistry, caseSensitive: Boolean)
  extends RuleExecutor[LogicalPlan] with HiveTypeCoercion {

  // TODO: pass this in as a parameter.
  val fixedPoint = FixedPoint(100)

  val batches: Seq[Batch] = Seq(
    Batch("MultiInstanceRelations", Once,
      NewRelationInstances),
    Batch("CaseInsensitiveAttributeReferences", Once,
      (if (caseSensitive) Nil else LowercaseAttributeReferences :: Nil) : _*),
    Batch("Resolution", fixedPoint,
      ResolveReferences ::
      ResolveRelations ::
      NewRelationInstances ::
      ImplicitGenerate ::
      StarExpansion ::
      ResolveFunctions ::
      GlobalAggregates ::
      typeCoercionRules :_*),
    Batch("AnalysisOperators", fixedPoint,
      EliminateAnalysisOperators)
  )

Analyzer里的一些对象解释：

FixedPoint：相当于迭代次数的上限。

  /** A strategy that runs until fix point or maxIterations times, whichever comes first. */
  case class FixedPoint(maxIterations: Int) extends Strategy

Batch: 批次，这个对象是由一系列Rule组成的，採用一个策略（策略事实上是迭代几次的别名吧，eg：Once）

  /** A batch of rules. */。
  protected case class Batch(name: String, strategy: Strategy, rules: Rule[TreeType]*)

Rule：理解为一种规则，这样的规则会应用到Logical Plan 从而将UnResolved 转变为Resolved

abstract class Rule[TreeType <: TreeNode[_]] extends Logging {

  /** Name for this rule, automatically inferred based on class name. */
  val ruleName: String = {
    val className = getClass.getName
    if (className endsWith "$") className.dropRight(1) else className
  }

  def apply(plan: TreeType): TreeType
}

Strategy：最大的运行次数，假设运行次数在最大迭代次数之前就达到了fix point，策略就会停止，不再应用了。

  /**
   * An execution strategy for rules that indicates the maximum number of executions. If the
   * execution reaches fix point (i.e. converge) before maxIterations, it will stop.
   */
  abstract class Strategy { def maxIterations: Int }

Analyzer解析主要是依据这些Batch里面定义的策略和Rule来对Unresolved的逻辑计划进行解析的。

这里Analyzer类本身并未定义运行的方法，而是要从它的父类RuleExecutor[LogicalPlan]寻找，Analyzer也实现了HiveTypeCosercion，这个类是參考Hive的类型自己主动兼容转换的原理。如图：

RuleExecutor：运行Rule的运行环境，它会将包括了一系列的Rule的Batch进行运行，这个过程都是串行的。

详细的运行方法定义在apply里：

能够看到这里是一个while循环，每一个batch下的rules都对当前的plan进行作用，这个过程是迭代的，直到达到Fix Point或者最大迭代次数。

 def apply(plan: TreeType): TreeType = {
    var curPlan = plan

    batches.foreach { batch =>
      val batchStartPlan = curPlan
      var iteration = 1
      var lastPlan = curPlan
      var continue = true

      // Run until fix point (or the max number of iterations as specified in the strategy.
      while (continue) {
        curPlan = batch.rules.foldLeft(curPlan) {
          case (plan, rule) =>
            val result = rule(plan) //这里将调用各个不同Rule的apply方法，将UnResolved Relations，Attrubute和Function进行Resolve
            if (!result.fastEquals(plan)) {
              logger.trace(
                s"""
                  |=== Applying Rule ${rule.ruleName} ===
                  |${sideBySide(plan.treeString, result.treeString).mkString("\n")}
                """.stripMargin)
            }

            result //返回作用后的result plan
        }
        iteration += 1
        if (iteration > batch.strategy.maxIterations) { //假设迭代次数已经大于该策略的最大迭代次数，就停止循环
          logger.info(s"Max iterations ($iteration) reached for batch ${batch.name}")
          continue = false
        }

        if (curPlan.fastEquals(lastPlan)) { //假设在多次迭代中不再变化，由于plan有个unique id，就停止循环。

logger.trace(s"Fixed point reached for batch ${batch.name} after $iteration iterations.")
          continue = false
        }
        lastPlan = curPlan
      }

      if (!batchStartPlan.fastEquals(curPlan)) {
        logger.debug(
          s"""
          |=== Result of Batch ${batch.name} ===
          |${sideBySide(plan.treeString, curPlan.treeString).mkString("\n")}
        """.stripMargin)
      } else {
        logger.trace(s"Batch ${batch.name} has no effect.")
      }
    }

    curPlan //返回Resolved的Logical Plan
  }

二、Rules介绍

    眼下Spark SQL 1.0.0的Rule都定义在了Analyzer.scala的内部类。

    在batches里面定义了4个Batch。

    MultiInstanceRelations、CaseInsensitiveAttributeReferences、Resolution、AnalysisOperators 四个。

    这4个Batch是将不同的Rule进行归类。每种类别採用不同的策略来进行Resolve。

    

2.1、MultiInstanceRelation

假设一个实例在Logical Plan里出现了多次，则会应用NewRelationInstances这儿Rule

 Batch("MultiInstanceRelations", Once,
      NewRelationInstances)

trait MultiInstanceRelation {
  def newInstance: this.type
}

object NewRelationInstances extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = {
    val localRelations = plan collect { case l: MultiInstanceRelation => l} //将logical plan应用partial function得到全部MultiInstanceRelation的plan的集合
    val multiAppearance = localRelations
      .groupBy(identity[MultiInstanceRelation]) //group by操作
      .filter { case (_, ls) => ls.size > 1 } //假设仅仅取size大于1的进行兴许操作
      .map(_._1)
      .toSet

    //更新plan，使得每一个实例的expId是唯一的。
    plan transform {
      case l: MultiInstanceRelation if multiAppearance contains l => l.newInstance
    }
  }
}

2.2、LowercaseAttributeReferences

相同是partital function，对当前plan应用，将全部匹配的如UnresolvedRelation的别名alise转换为小写，将Subquery的别名也转换为小写。

总结：这是一个使属性名大写和小写不敏感的Rule，由于它将全部属性都to lower case了。

  object LowercaseAttributeReferences extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transform {
      case UnresolvedRelation(databaseName, name, alias) =>
        UnresolvedRelation(databaseName, name, alias.map(_.toLowerCase))
      case Subquery(alias, child) => Subquery(alias.toLowerCase, child)
      case q: LogicalPlan => q transformExpressions {
        case s: Star => s.copy(table = s.table.map(_.toLowerCase))
        case UnresolvedAttribute(name) => UnresolvedAttribute(name.toLowerCase)
        case Alias(c, name) => Alias(c, name.toLowerCase)()
        case GetField(c, name) => GetField(c, name.toLowerCase)
      }
    }
  }

2.3、ResolveReferences

将Sql parser解析出来的UnresolvedAttribute全部都转为相应的实际的catalyst.expressions.AttributeReference AttributeReferences

这里调用了logical plan 的resolve方法。将属性转为NamedExepression。

  object ResolveReferences extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {
      case q: LogicalPlan if q.childrenResolved =>
        logger.trace(s"Attempting to resolve ${q.simpleString}")
        q transformExpressions {
          case u @ UnresolvedAttribute(name) =>
            // Leave unchanged if resolution fails.  Hopefully will be resolved next round.
            val result = q.resolve(name).getOrElse(u)//转化为NamedExpression
            logger.debug(s"Resolving $u to $result")
            result
        }
    }
  }

2.4、 ResolveRelations

这个比較好理解，还记得前面Sql parser吗。比方select * from src，这个src表parse后就是一个UnresolvedRelation节点。

这一步ResolveRelations调用了catalog这个对象。

Catalog对象里面维护了一个tableName,Logical Plan的HashMap结果。

通过这个Catalog文件夹来寻找当前表的结构。从而从中解析出这个表的字段，如UnResolvedRelations 会得到一个tableWithQualifiers。（即表和字段） 

这也解释了为什么流程图那，我会画一个catalog在上面。由于它是Analyzer工作时须要的meta data。

object ResolveRelations extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transform {
      case UnresolvedRelation(databaseName, name, alias) =>
        catalog.lookupRelation(databaseName, name, alias)
    }
  }

2.5、ImplicitGenerate

假设在select语句里仅仅有一个表达式，并且这个表达式是一个Generator（Generator是一个1条记录生成到N条记录的映射）

当在解析逻辑计划时，遇到Project节点的时候，就能够将它转换为Generate类（Generate类是将输入流应用一个函数。从而生成一个新的流）。

  object ImplicitGenerate extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transform {
      case Project(Seq(Alias(g: Generator, _)), child) =>
        Generate(g, join = false, outer = false, None, child)
    }
  }

2.6 StarExpansion

在Project操作符里。假设是*符号，即select * 语句。能够将全部的references都展开，即将select * 中的*展开成实际的字段。

  object StarExpansion extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transform {
      // Wait until children are resolved
      case p: LogicalPlan if !p.childrenResolved => p
      // If the projection list contains Stars, expand it.
      case p @ Project(projectList, child) if containsStar(projectList) =>
        Project(
          projectList.flatMap {
            case s: Star => s.expand(child.output) //展开，将输入的Attributeexpand(input: Seq[Attribute]) 转化为Seq[NamedExpression]
            case o => o :: Nil
          },
          child)
      case t: ScriptTransformation if containsStar(t.input) =>
        t.copy(
          input = t.input.flatMap {
            case s: Star => s.expand(t.child.output)
            case o => o :: Nil
          }
        )
      // If the aggregate function argument contains Stars, expand it.
      case a: Aggregate if containsStar(a.aggregateExpressions) =>
        a.copy(
          aggregateExpressions = a.aggregateExpressions.flatMap {
            case s: Star => s.expand(a.child.output)
            case o => o :: Nil
          }
        )
    }
    /**
     * Returns true if `exprs` contains a [[Star]].
     */
    protected def containsStar(exprs: Seq[Expression]): Boolean =
      exprs.collect { case _: Star => true }.nonEmpty
  }
}

2.7 ResolveFunctions

这个和ResolveReferences几乎相同，这里主要是对udf进行resolve。

将这些UDF都在FunctionRegistry里进行查找。

  object ResolveFunctions extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transform {
      case q: LogicalPlan =>
        q transformExpressions {
          case u @ UnresolvedFunction(name, children) if u.childrenResolved =>
            registry.lookupFunction(name, children) //看是否注冊了当前udf
        }
    }
  }

2.8 GlobalAggregates

全局的聚合，假设遇到了Project就返回一个Aggregate.

  object GlobalAggregates extends Rule[LogicalPlan] {
    def apply(plan: LogicalPlan): LogicalPlan = plan transform {
      case Project(projectList, child) if containsAggregates(projectList) =>
        Aggregate(Nil, projectList, child)
    }

    def containsAggregates(exprs: Seq[Expression]): Boolean = {
      exprs.foreach(_.foreach {
        case agg: AggregateExpression => return true
        case _ =>
      })
      false
    }
  }

2.9 typeCoercionRules

这个是Hive里的兼容SQL语法，比方将String和Int互相转换，不须要显示的调用cast xxx  as yyy了。如StringToIntegerCasts。

  val typeCoercionRules =
    PropagateTypes ::
    ConvertNaNs ::
    WidenTypes ::
    PromoteStrings ::
    BooleanComparisons ::
    BooleanCasts ::
    StringToIntegralCasts ::
    FunctionArgumentConversion ::
    CastNulls ::
    Nil

2.10 EliminateAnalysisOperators

将分析的操作符移除。这里仅支持2种，一种是Subquery须要移除，一种是LowerCaseSchema。这些节点都会从Logical Plan里移除。

object EliminateAnalysisOperators extends Rule[LogicalPlan] {
  def apply(plan: LogicalPlan): LogicalPlan = plan transform {
    case Subquery(_, child) => child //遇到Subquery,不反悔本身，返回它的Child，即删除了该元素
    case LowerCaseSchema(child) => child
  }
}

三、实践

  补充昨天DEBUG的一个样例。这个样例证实了如何将上面的规则应用到Unresolved Logical Plan：

  当传递sql语句的时候，的确调用了ResolveReferences将mobile解析成NamedExpression。

  能够对比这看运行流程，左边是Unresolved Logical Plan，右边是Resoveld Logical Plan。

  先是运行了Batch Resolution，eg: 调用ResovelRalation这个RUle来使 Unresovled Relation 转化为 SparkLogicalPlan并通过Catalog找到了其对于的字段属性。

  然后运行了Batch Analysis Operator。eg：调用EliminateAnalysisOperators来将SubQuery给remove掉了。

  可能格式显示的不太好，能够向右边拖动下滚动轴看下结果。 ：） 

  

val exec = sqlContext.sql("select mobile as mb, sid as id, mobile*2 multi2mobile, count(1) times from (select * from temp_shengli_mobile)a where pfrom_id=0.0 group by mobile, sid,  mobile*2")
14/07/21 18:23:32 DEBUG SparkILoop$SparkILoopInterpreter: Invoking: public static java.lang.String $line47.$eval.$print()
14/07/21 18:23:33 INFO Analyzer: Max iterations (2) reached for batch MultiInstanceRelations
14/07/21 18:23:33 INFO Analyzer: Max iterations (2) reached for batch CaseInsensitiveAttributeReferences
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'pfrom_id to pfrom_id#5
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'sid to sid#1
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'sid to sid#1
14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2
14/07/21 18:23:33 DEBUG Analyzer:
=== Result of Batch Resolution ===
!Aggregate ['mobile,'sid,('mobile * 2) AS c2#27], ['mobile AS mb#23,'sid AS id#24,('mobile * 2) AS multi2mobile#25,COUNT(1) AS times#26L]   Aggregate [mobile#2,sid#1,(CAST(mobile#2, DoubleType) * CAST(2, DoubleType)) AS c2#27], [mobile#2 AS mb#23,sid#1 AS id#24,(CAST(mobile#2, DoubleType) * CAST(2, DoubleType)) AS multi2mobile#25,COUNT(1) AS times#26L]
! Filter ('pfrom_id = 0.0)                                                                                                                   Filter (CAST(pfrom_id#5, DoubleType) = 0.0)
   Subquery a                                                                                                                                 Subquery a
!   Project [*]                                                                                                                                Project [data_date#0,sid#1,mobile#2,pverify_type#3,create_time#4,pfrom_id#5,p_status#6,pvalidate_time#7,feffect_time#8,plastupdate_ip#9,update_time#10,status#11,preserve_int#12]
!    UnresolvedRelation None, temp_shengli_mobile, None                                                                                         Subquery temp_shengli_mobile
!                                                                                                                                                SparkLogicalPlan (ExistingRdd [data_date#0,sid#1,mobile#2,pverify_type#3,create_time#4,pfrom_id#5,p_status#6,pvalidate_time#7,feffect_time#8,plastupdate_ip#9,update_time#10,status#11,preserve_int#12], MapPartitionsRDD[4] at mapPartitions at basicOperators.scala:174)

14/07/21 18:23:33 DEBUG Analyzer:
=== Result of Batch AnalysisOperators ===
!Aggregate ['mobile,'sid,('mobile * 2) AS c2#27], ['mobile AS mb#23,'sid AS id#24,('mobile * 2) AS multi2mobile#25,COUNT(1) AS times#26L]   Aggregate [mobile#2,sid#1,(CAST(mobile#2, DoubleType) * CAST(2, DoubleType)) AS c2#27], [mobile#2 AS mb#23,sid#1 AS id#24,(CAST(mobile#2, DoubleType) * CAST(2, DoubleType)) AS multi2mobile#25,COUNT(1) AS times#26L]
! Filter ('pfrom_id = 0.0)                                                                                                                   Filter (CAST(pfrom_id#5, DoubleType) = 0.0)
!  Subquery a                                                                                                                                 Project [data_date#0,sid#1,mobile#2,pverify_type#3,create_time#4,pfrom_id#5,p_status#6,pvalidate_time#7,feffect_time#8,plastupdate_ip#9,update_time#10,status#11,preserve_int#12]
!   Project [*]                                                                                                                                SparkLogicalPlan (ExistingRdd [data_date#0,sid#1,mobile#2,pverify_type#3,create_time#4,pfrom_id#5,p_status#6,pvalidate_time#7,feffect_time#8,plastupdate_ip#9,update_time#10,status#11,preserve_int#12], MapPartitionsRDD[4] at mapPartitions at basicOperators.scala:174)
!    UnresolvedRelation None, temp_shengli_mobile, None
        

四、总结

    本文从源代码角度分析了Analyzer在对Sql Parser解析出的UnResolve Logical Plan 进行analyze的过程中，所运行的流程。

    流程是实例化一个SimpleAnalyzer，定义一些Batch，然后遍历这些Batch在RuleExecutor的环境下。运行Batch里面的Rules。每一个Rule会对Unresolved Logical Plan进行Resolve，有些可能不能一次解析出。须要多次迭代，直到达到max迭代次数或者达到fix point。这里Rule里比較经常使用的就是ResolveReferences、ResolveRelations、StarExpansion、GlobalAggregates、typeCoercionRules和EliminateAnalysisOperators。

——EOF——

原创文章，转载请注明：

转载自：OopsOutOfMemory盛利的Blog，作者： OopsOutOfMemory

本文链接地址：http://blog.csdn.net/oopsoom/article/details/38025185

注：本文基于署名-非商业性使用-禁止演绎 2.5 中国大陆(CC BY-NC-ND 2.5 CN)协议，欢迎转载、转发和评论，可是请保留本文作者署名和文章链接。如若须要用于商业目的或者与授权方面的协商，请联系我。