Spark SQL Catalyst源码分析之Analyzer

 /** Spark SQL源码分析系列文章*/

    前面几篇文章讲解了Spark SQL的核心执行流程和Spark SQL的Catalyst框架的Sql Parser是怎样接受用户输入sql，经过解析生成Unresolved Logical Plan的。我们记得Spark SQL的执行流程中另一个核心的组件式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。

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].

1. class Analyzer(catalog: Catalog, registry: FunctionRegistry, caseSensitive: Boolean)

2. extends RuleExecutor[LogicalPlan] with HiveTypeCoercion {

3.  

4. // TODO: pass this in as a parameter.

5. val fixedPoint = FixedPoint(100)

6.  

7. val batches: Seq[Batch] = Seq(

8. Batch("MultiInstanceRelations", Once,

9. NewRelationInstances),

10. Batch("CaseInsensitiveAttributeReferences", Once,

11. (if (caseSensitive) Nil else LowercaseAttributeReferences :: Nil) : _*),

12. Batch("Resolution", fixedPoint,

13. ResolveReferences ::

14. ResolveRelations ::

15. NewRelationInstances ::

16. ImplicitGenerate ::

17. StarExpansion ::

18. ResolveFunctions ::

19. GlobalAggregates ::

20. typeCoercionRules :_*),

21. Batch("AnalysisOperators", fixedPoint,

22. EliminateAnalysisOperators)

23. )

    Analyzer里的一些对象解释：

    FixedPoint：相当于迭代次数的上限。

1. /** A strategy that runs until fix point or maxIterations times, whichever comes first. */

2. case class FixedPoint(maxIterations: Int) extends Strategy

    Batch: 批次，这个对象是由一系列Rule组成的，采用一个策略（策略其实是迭代几次的别名吧，eg：Once）

1. /** A batch of rules. */，

2. protected case class Batch(name: String, strategy: Strategy, rules: Rule[TreeType]*)

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

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

2.  

3. /** Name for this rule, automatically inferred based on class name. */

4. val ruleName: String = {

5. val className = getClass.getName

6. if (className endsWith "$") className.dropRight(1) else className

7. }

8.  

9. def apply(plan: TreeType): TreeType

10. }

   Strategy：最大的执行次数，如果执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。

1. /**

2. * An execution strategy for rules that indicates the maximum number of executions. If the

3. * execution reaches fix point (i.e. converge) before maxIterations, it will stop.

4. */

5. abstract class Strategy { def maxIterations: Int }

   Analyzer解析主要是根据这些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类本身并没有定义执行的方法，而是要从它的父类RuleExecutor[LogicalPlan]寻找，Analyzer也实现了HiveTypeCosercion，这个类是参考Hive的类型自动兼容转换的原理。如图：

    RuleExecutor：执行Rule的执行环境，它会将包含了一系列的Rule的Batch进行执行，这个过程都是串行的。

    具体的执行方法定义在apply里：

    可以看到这里是一个while循环，每个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。

1. def apply(plan: TreeType): TreeType = {

2. var curPlan = plan

3.  

4. batches.foreach { batch =>

5. val batchStartPlan = curPlan

6. var iteration = 1

7. var lastPlan = curPlan

8. var continue = true

9.  

10. // Run until fix point (or the max number of iterations as specified in the strategy.

11. while (continue) {

12. curPlan = batch.rules.foldLeft(curPlan) {

13. case (plan, rule) =>

14. 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

15. if (!result.fastEquals(plan)) {

16. logger.trace(

17. s"""

18. |=== Applying Rule ${rule.ruleName} ===

19. |${sideBySide(plan.treeString, result.treeString).mkString("n")}

20. """.stripMargin)

21. }

22.  

23. result //返回作用后的result plan

24. }

25. iteration += 1

26. if (iteration > batch.strategy.maxIterations) { //如果迭代次数已经大于该策略的最大迭代次数，就停止循环

27. logger.info(s"Max iterations ($iteration) reached for batch ${batch.name}")

28. continue = false

29. }

30.  

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

32. logger.trace(s"Fixed point reached for batch ${batch.name} after $iteration iterations.")

33. continue = false

34. }

35. lastPlan = curPlan

36. }

37.  

38. if (!batchStartPlan.fastEquals(curPlan)) {

39. logger.debug(

40. s"""

41. |=== Result of Batch ${batch.name} ===

42. |${sideBySide(plan.treeString, curPlan.treeString).mkString("n")}

43. """.stripMargin)

44. } else {

45. logger.trace(s"Batch ${batch.name} has no effect.")

46. }

47. }

48.  

49. curPlan //返回Resolved的Logical Plan

50. }

二、Rules介绍

    目前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，每种类别采用不同的策略来进行Resolve。

2.1、MultiInstanceRelation 

如果一个实例在Logical Plan里出现了多次，则会应用NewRelationInstances这儿Rule

1. Batch("MultiInstanceRelations", Once,

2. NewRelationInstances)

1. trait MultiInstanceRelation {

2.   def newInstance: this.type

3. }

1. object NewRelationInstances extends Rule[LogicalPlan] {

2.   def apply(plan: LogicalPlan): LogicalPlan = {

3.     val localRelations = plan collect { case l: MultiInstanceRelation => l} //将logical plan应用partial function得到所有MultiInstanceRelation的plan的集合

4.     val multiAppearance = localRelations

5.       .groupBy(identity[MultiInstanceRelation]) //group by操作

6.       .filter { case (_, ls) => ls.size > 1 } //如果只取size大于1的进行后续操作

7.       .map(_._1)

8.       .toSet

9.  

10. //更新plan，使得每个实例的expId是唯一的。

11.     plan transform {

12.       case l: MultiInstanceRelation if multiAppearance contains l => l.newInstance

13.     }

14.   }

15. }

2.2、LowercaseAttributeReferences

同样是partital function，对当前plan应用，将所有匹配的如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小写不敏感的Rule，因为它将所有属性都to lower case了。

1. object LowercaseAttributeReferences extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3. case UnresolvedRelation(databaseName, name, alias) =>

4. UnresolvedRelation(databaseName, name, alias.map(_.toLowerCase))

5. case Subquery(alias, child) => Subquery(alias.toLowerCase, child)

6. case q: LogicalPlan => q transformExpressions {

7. case s: Star => s.copy(table = s.table.map(_.toLowerCase))

8. case UnresolvedAttribute(name) => UnresolvedAttribute(name.toLowerCase)

9. case Alias(c, name) => Alias(c, name.toLowerCase)()

10. case GetField(c, name) => GetField(c, name.toLowerCase)

11. }

12. }

13. }

2.3、ResolveReferences

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

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

1. object ResolveReferences extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {

3. case q: LogicalPlan if q.childrenResolved =>

4. logger.trace(s"Attempting to resolve ${q.simpleString}")

5. q transformExpressions {

6. case u @ UnresolvedAttribute(name) =>

7. // Leave unchanged if resolution fails. Hopefully will be resolved next round.

8. val result = q.resolve(name).getOrElse(u)//转化为NamedExpression

9. logger.debug(s"Resolving $u to $result")

10. result

11. }

12. }

13. }

2.4、 ResolveRelations

这个比较好理解，还记得前面Sql parser吗，比如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目录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，因为它是Analyzer工作时需要的meta data。

1. object ResolveRelations extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3. case UnresolvedRelation(databaseName, name, alias) =>

4. catalog.lookupRelation(databaseName, name, alias)

5. }

6. }

2.5、ImplicitGenerate

如果在select语句里只有一个表达式，而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，就可以将它转换为Generate类（Generate类是将输入流应用一个函数，从而生成一个新的流）。

1. object ImplicitGenerate extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3. case Project(Seq(Alias(g: Generator, _)), child) =>

4. Generate(g, join = false, outer = false, None, child)

5. }

6. }

2.6 StarExpansion

在Project操作符里，如果是*符号，即select * 语句，可以将所有的references都展开，即将select * 中的*展开成实际的字段。

1.   object StarExpansion extends Rule[LogicalPlan] {

2.     def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3.       // Wait until children are resolved

4.       case p: LogicalPlan if !p.childrenResolved => p

5.       // If the projection list contains Stars, expand it.

6.       case p @ Project(projectList, child) if containsStar(projectList) =>

7.         Project(

8.           projectList.flatMap {

9.             case s: Star => s.expand(child.output) //展开，将输入的Attributeexpand(input: Seq[Attribute]) 转化为Seq[NamedExpression]

10.             case o => o :: Nil

11.           },

12.           child)

13.       case t: ScriptTransformation if containsStar(t.input) =>

14.         t.copy(

15.           input = t.input.flatMap {

16.             case s: Star => s.expand(t.child.output)

17.             case o => o :: Nil

18.           }

19.         )

20.       // If the aggregate function argument contains Stars, expand it.

21.       case a: Aggregate if containsStar(a.aggregateExpressions) =>

22.         a.copy(

23.           aggregateExpressions = a.aggregateExpressions.flatMap {

24.             case s: Star => s.expand(a.child.output)

25.             case o => o :: Nil

26.           }

27.         )

28.     }

29.     /**

30.      * Returns true if `exprs` contains a [[Star]].

31.      */

32.     protected def containsStar(exprs: Seq[Expression]): Boolean =

33.       exprs.collect { case _: Star => true }.nonEmpty

34.   }

35. }

2.7 ResolveFunctions

这个和ResolveReferences差不多，这里主要是对udf进行resolve。

将这些UDF都在FunctionRegistry里进行查找。

1. object ResolveFunctions extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3. case q: LogicalPlan =>

4. q transformExpressions {

5. case u @ UnresolvedFunction(name, children) if u.childrenResolved =>

6. registry.lookupFunction(name, children) //看是否注册了当前udf

7. }

8. }

9. }

2.8 GlobalAggregates

全局的聚合，如果遇到了Project就返回一个Aggregate.

1. object GlobalAggregates extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3. case Project(projectList, child) if containsAggregates(projectList) =>

4. Aggregate(Nil, projectList, child)

5. }

6.  

7. def containsAggregates(exprs: Seq[Expression]): Boolean = {

8. exprs.foreach(_.foreach {

9. case agg: AggregateExpression => return true

10. case _ =>

11. })

12. false

13. }

14. }

2.9 typeCoercionRules

这个是Hive里的兼容SQL语法，比如将String和Int互相转换，不需要显示的调用cast xxx  as yyy了。如StringToIntegerCasts。

1. val typeCoercionRules =

2. PropagateTypes ::

3. ConvertNaNs ::

4. WidenTypes ::

5. PromoteStrings ::

6. BooleanComparisons ::

7. BooleanCasts ::

8. StringToIntegralCasts ::

9. FunctionArgumentConversion ::

10. CastNulls ::

11. Nil

2.10 EliminateAnalysisOperators

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

1. object EliminateAnalysisOperators extends Rule[LogicalPlan] {

2. def apply(plan: LogicalPlan): LogicalPlan = plan transform {

3. case Subquery(_, child) => child //遇到Subquery,不反悔本身，返回它的Child，即删除了该元素

4. case LowerCaseSchema(child) => child

5. }

6. }

三、实践

  补充昨天DEBUG的一个例子，这个例子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。

  可以对照这看执行流程，左边是Unresolved Logical Plan，右边是Resoveld Logical Plan。

  先是执行了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。

  然后执行了Batch Analysis Operator。eg：调用EliminateAnalysisOperators来将SubQuery给remove掉了。

  可能格式显示的不太好，可以向右边拖动下滚动轴看下结果。 ：） 

1. 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")

2. 14/07/21 18:23:32 DEBUG SparkILoop$SparkILoopInterpreter: Invoking: public static java.lang.String $line47.$eval.$print()

3. 14/07/21 18:23:33 INFO Analyzer: Max iterations (2) reached for batch MultiInstanceRelations

4. 14/07/21 18:23:33 INFO Analyzer: Max iterations (2) reached for batch CaseInsensitiveAttributeReferences

5. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'pfrom_id to pfrom_id#5

6. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2

7. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'sid to sid#1

8. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2

9. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2

10. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'sid to sid#1

11. 14/07/21 18:23:33 DEBUG Analyzer$ResolveReferences$: Resolving 'mobile to mobile#2

12. 14/07/21 18:23:33 DEBUG Analyzer:

13. === Result of Batch Resolution ===

14. !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]

15. ! Filter ('pfrom_id = 0.0) Filter (CAST(pfrom_id#5, DoubleType) = 0.0)

16. Subquery a Subquery a

17. ! 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]

18. ! UnresolvedRelation None, temp_shengli_mobile, None Subquery temp_shengli_mobile

19. ! 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)

20.  

21. 14/07/21 18:23:33 DEBUG Analyzer:

22. === Result of Batch AnalysisOperators ===

23. !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]

24. ! Filter ('pfrom_id = 0.0) Filter (CAST(pfrom_id#5, DoubleType) = 0.0)

25. ! 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]

26. ! 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)

27. ! 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，所执行的流程。

    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，执行Batch里面的Rules，每个Rule会对Unresolved Logical Plan进行Resolve，有些可能不能一次解析出，需要多次迭代，直到达到max迭代次数或者达到fix point。这里Rule里比较常用的就是ResolveReferences、ResolveRelations、StarExpansion、GlobalAggregates、typeCoercionRules和EliminateAnalysisOperators。

最后

以上就是老实发夹为你收集整理的Spark SQL Catalyst源码分析之Analyzer的全部内容，希望文章能够帮你解决Spark SQL Catalyst源码分析之Analyzer所遇到的程序开发问题。

如果觉得靠谱客网站的内容还不错，欢迎将靠谱客网站推荐给程序员好友。