Spark SQL发展历程 spark sql原理

转载

技术领航博主 2023-06-05 10:44:39

文章标签 Spark SQL发展历程 SQL ide sql 文章分类 Spark 大数据

一. 整体架构

总结为如下图：

Dataframe本质是数据 + 数据的描述信息（结构元信息）

所有的上述SQL及dataframe操作最终都通过Catalyst翻译成spark程序RDD操作代码

sparkSQL前身是shark，大量依赖Hive项目的jar包与功能，但在上面的扩展越来越难，因此出现了SparkSQL，它重写了分析器，执行器 脱离了对Hive项目的大部分依赖，基本可以独立去运行，只用到两个地方：

1.借用了hive的词汇分析的jar即HiveQL解析器

2.借用了hive的metastore和数据访问API即hiveCatalog

也就是说上图的左半部分的操作全部用的是sparkSQL本身自带的内置SQL解析器解析SQL进行翻译，用到内置元数据信息(比如结构化文件中自带的结构元信息，RDD的schema中的结构元信息)

右半部分则是走的Hive的HQL解析器，还有Hive元数据信息

因此左右两边的API调用的底层类会有不同

SQLContext使用：

简单的解析器（scala语言写的sql解析器）【比如：1.在半结构化的文件里面使用sql查询时，是用这个解析器解析的，2.访问（半）结构化文件的时候，通过sqlContext使用schema，类生成Dataframe，然后dataframe注册为表时，registAsTmpTable 然后从这个表里面进行查询时，即使用的简单的解析器，一些hive语法应该是不支持的，有待验证）】

simpleCatalog【此对象中存放关系（表），比如我们指定的schema信息，类的信息，都是关系信息】

HiveContext使用：

HiveQL解析器【支持hive的hql语法，如只有通过HiveContext生成的dataframe才能调用saveAsTable操作】

hiveCatalog（存放数据库和表的元数据信息）

Sparksql的解析与Hiveql的解析的执行流程:

Spark SQL发展历程 spark sql原理_ide_02

一个Sql语句转化为实际可执行的Spark的RDD模型需要经过以下几个步骤：

Spark SQL发展历程 spark sql原理_Spark SQL发展历程_03

主要介绍下Spark-SQL里面的主要类成员：

1.2 SQLContext

SQL上下文环境，它保存了QueryExecution中所需要的几个类：

1.2.1 Catalog

一个存储<tableName,logicalPlan>的map结构，查找关系的目录，注册表，注销表，查询表和逻辑计划关系的类

1. @transient  
2. protected[sql] lazy val catalog: Catalog = new SimpleCatalog(conf)  
3. class SimpleCatalog(val conf: CatalystConf) extends Catalog {  
4. new mutable.HashMap[String, LogicalPlan]()  
5.   override def registerTable(  
6.       tableIdentifier: Seq[String],  
7.       plan: LogicalPlan): Unit = {  
8. //转化大小写  
9.     val tableIdent = processTableIdentifier(tableIdentifier)  
10.     tables += ((getDbTableName(tableIdent), plan))  
11.   }  
12.   override def unregisterTable(tableIdentifier: Seq[String]): Unit = {  
13.     val tableIdent = processTableIdentifier(tableIdentifier)  
14.     tables -= getDbTableName(tableIdent)  
15.   }  
16.   override def unregisterAllTables(): Unit = {  
17.     tables.clear()  
18.   }  
19.   override def tableExists(tableIdentifier: Seq[String]): Boolean = {  
20.     val tableIdent = processTableIdentifier(tableIdentifier)  
21.     tables.get(getDbTableName(tableIdent)) match {  
22. case Some(_) => true  
23. case None => false  
24.     }  
25.   }  
26.   override def lookupRelation(  
27.       tableIdentifier: Seq[String],  
28.       alias: Option[String] = None): LogicalPlan = {  
29.     val tableIdent = processTableIdentifier(tableIdentifier)  
30.     val tableFullName = getDbTableName(tableIdent)  
31. //  val tables = new mutable.HashMap[String, LogicalPlan](),根据表名获取logicalplan  
32. "Table Not Found: $tableFullName"))  
33.     val tableWithQualifiers = Subquery(tableIdent.last, table)  
34. // If an alias was specified by the lookup, wrap the plan in a subquery so that attributes are  
35. // properly qualified with this alias.  
36.     alias.map(a => Subquery(a, tableWithQualifiers)).getOrElse(tableWithQualifiers)  
37.   }  
38.   override def getTables(databaseName: Option[String]): Seq[(String, Boolean)] = {  
39.     tables.map {  
40. case (name, _) => (name, true)  
41.     }.toSeq  
42.   }  
43.   override def refreshTable(databaseName: String, tableName: String): Unit = {  
44. throw new UnsupportedOperationException  
45.   }  
46. }

1.2.2 SparkSQLParser

将Sql语句解析成语法树，返回一个Logical Plan。它首先拆分不同的SQL（将其分类），然后利用fallback解析。

1. /**
2.  * The top level Spark SQL parser. This parser recognizes syntaxes that are available for all SQL
3.  * dialects supported by Spark SQL, and delegates all the other syntaxes to the `fallback` parser.
4.  *
5.  * @param fallback A function that parses an input string to a logical plan
6.  */  
7. private[sql] class SparkSQLParser(fallback: String => LogicalPlan) extends AbstractSparkSQLParser {  
8. protected val AS = Keyword("AS")  
9. protected val CACHE = Keyword("CACHE")  
10. protected val CLEAR = Keyword("CLEAR")  
11. protected val IN = Keyword("IN")  
12. protected val LAZY = Keyword("LAZY")  
13. protected val SET = Keyword("SET")  
14. protected val SHOW = Keyword("SHOW")  
15. protected val TABLE = Keyword("TABLE")  
16. protected val TABLES = Keyword("TABLES")  
17. protected val UNCACHE = Keyword("UNCACHE")  
18. override protected lazy val start: Parser[LogicalPlan] = cache | uncache | set | show | others  
19. private lazy val cache: Parser[LogicalPlan] =  
20.   CACHE ~> LAZY.? ~ (TABLE ~> ident) ~ (AS ~> restInput).? ^^ {  
21. case isLazy ~ tableName ~ plan =>  
22.       CacheTableCommand(tableName, plan.map(fallback), isLazy.isDefined)  
23.   }  
24. private lazy val uncache: Parser[LogicalPlan] =  
25.   ( UNCACHE ~ TABLE ~> ident ^^ {  
26. case tableName => UncacheTableCommand(tableName)  
27.     }  
28.   | CLEAR ~ CACHE ^^^ ClearCacheCommand  
29.   )  
30. private lazy val set: Parser[LogicalPlan] =  
31.   SET ~> restInput ^^ {  
32. case input => SetCommandParser(input)  
33.   }  
34. private lazy val show: Parser[LogicalPlan] =  
35.   SHOW ~> TABLES ~ (IN ~> ident).? ^^ {  
36. case _ ~ dbName => ShowTablesCommand(dbName)  
37.   }  
38. private lazy val others: Parser[LogicalPlan] =  
39.   wholeInput ^^ {  
40. case input => fallback(input)  
41.   }  
42. }

1.2.3 Analyzer

语法分析器，Analyzer会使用Catalog和FunctionRegistry将UnresolvedAttribute和UnresolvedRelation转换为catalyst里全类型的对象。例如将

'UnresolvedRelation[test], None

转化为

Relation[id#0L,dev_id#1,dev_chnnum#2L,dev_name#3,dev_chnname#4,car_num#5,car_numtype#6,car_numcolor#7,car_speed#8,car_type#9,car_color#10,car_length#11L,car_direct#12,car_way_code#13,cap_time#14L,cap_date#15L,inf_note#16,max_speed#17,min_speed#18,car_img_url#19,car_img1_url#20,car_img2_url#21,car_img3_url#22,car_img4_url#23,car_img5_url#24,rec_stat#25,dev_chnid#26,car_img_count#27,save_flag#28,dc_cleanflag#29,pic_id#30,car_img_plate_top#31L,car_img_plate_left#32L,car_img_plate_bottom#33L,car_img_plate_right#34L,car_brand#35L,issafetybelt#36,isvisor#37,bind_stat#38,car_num_pic#39,combined_pic_url#40,verify_memo#41,rec_stat_tmp#42]org.apache.spark.sql.parquet.ParquetRelation2@2a400010

1. class Analyzer(  
2.     catalog: Catalog,  
3.     registry: FunctionRegistry,  
4.     conf: CatalystConf,  
5. 100)  
6. extends RuleExecutor[LogicalPlan] with HiveTypeCoercion with CheckAnalysis {  
7. ……  
8. }

1.2.4 Optimizer

优化器，将Logical Plan进一步进行优化

1. object DefaultOptimizer extends Optimizer {  
2.   val batches =  
3. // SubQueries are only needed for analysis and can be removed before execution.  
4. "Remove SubQueries", FixedPoint(100),  
5.       EliminateSubQueries) ::  
6. "Operator Reordering", FixedPoint(100),  
7.       UnionPushdown,  
8.       CombineFilters,  
9.       PushPredicateThroughProject,  
10.       PushPredicateThroughJoin,  
11.       PushPredicateThroughGenerate,  
12.       ColumnPruning,  
13.       ProjectCollapsing,  
14.       CombineLimits) ::  
15. "ConstantFolding", FixedPoint(100),  
16.     NullPropagation,  
17.       OptimizeIn,  
18.       ConstantFolding,  
19.       LikeSimplification,  
20.       BooleanSimplification,  
21.       SimplifyFilters,  
22.       SimplifyCasts,  
23.       SimplifyCaseConversionExpressions) ::  
24. "Decimal Optimizations", FixedPoint(100),  
25.       DecimalAggregates) ::  
26. "LocalRelation", FixedPoint(100),  
27.       ConvertToLocalRelation) :: Nil  
28. }

例如:

CombineFilters:递归合并两个相邻的filter。例如：将

Filter(a>1)

Filter(b>1)

Project……

转化为

Filter(a>1) AND Filter(b>1)

Project……

CombineLimits:合并两个相邻的limit。例如：将select * from (select * from c_picrecord limit 100)a limit 10

优化为：

Limit if ((100 < 10)) 100 else 10
Relation[id#0L,dev_id#1,dev_chnnum#2L,de……

1.2.5 SparkPlanner

将LogicalPlan转化为SparkPlan

1. protected[sql] class SparkPlanner extends SparkStrategies {  
2.   val sparkContext: SparkContext = self.sparkContext  
3.   val sqlContext: SQLContext = self  
4.   def codegenEnabled: Boolean = self.conf.codegenEnabled  
5.   def unsafeEnabled: Boolean = self.conf.unsafeEnabled  
6.   def numPartitions: Int = self.conf.numShufflePartitions  
7.   def strategies: Seq[Strategy] =  
8.     experimental.extraStrategies ++ (  
9.     DataSourceStrategy ::  
10.     DDLStrategy ::  
11.     TakeOrdered ::  
12.     HashAggregation ::  
13.     LeftSemiJoin ::  
14.     HashJoin ::  
15.     InMemoryScans ::  
16.       ParquetOperations ::  
17.       BasicOperators ::  
18.       CartesianProduct ::  
19.       BroadcastNestedLoopJoin :: Nil)  
20.  }

比方说：

Subquery test

通过DataSourceStrategy中的

[java] view plain copy
 


// Scanning non-partitioned HadoopFsRelation  
 
case PhysicalOperation(projectList, filters, l @ LogicalRelation(t: HadoopFsRelation)) =>

将其转化为PhysicalRDD

1.2.6 PrepareForExecution

在SparkPlan中插入Shuffle的操作，如果前后2个SparkPlan的outputPartitioning不一样的话，则中间需要插入Shuffle的动作，比分说聚合函数，先局部聚合，然后全局聚合，局部聚合和全局聚合的分区规则是不一样的，中间需要进行一次Shuffle。

[java] view plain copy
 


/** 
 
 * Prepares a planned SparkPlan for execution by inserting shuffle operations as needed. 
 
 */  
 
@transient  
 
protected[sql] val prepareForExecution = new RuleExecutor[SparkPlan] {  
 
  val batches =  
 
    Batch("Add exchange", Once, EnsureRequirements(self)) :: Nil  
 
}  
 

例如

GeneratedAggregate false,[Coalesce(SUM(PartialCount#44L),0) AS count#43L], false

GeneratedAggregatetrue, [COUNT(1) AS PartialCount#44L], false

PhysicalRDDMapPartitionsRDD[1]

经过PrepareForExecution，转化为

GeneratedAggregate false,[Coalesce(SUM(PartialCount#44L),0) AS count#43L], false

Exchange SinglePartition

GeneratedAggregate true, [COUNT(1) AS PartialCount#44L], false

PhysicalRDDMapPartitionsRDD[1]

1.3 QueryExecution

SQL语句执行环境

1. protected[sql] class QueryExecution(val logical: LogicalPlan) {//logical包含了Aggregate(groupingExprs, aggregates, df.logicalPlan)  
2.   def assertAnalyzed(): Unit = analyzer.checkAnalysis(analyzed)  
3.   lazy val analyzed: LogicalPlan = analyzer.execute(logical)  
4.   lazy val withCachedData: LogicalPlan = {  
5.     assertAnalyzed()  
6.     cacheManager.useCachedData(analyzed)  
7.   }  
8. //优化过的LogicalPlan  
9. // TODO: Don't just pick the first one...  
10.   lazy val sparkPlan: SparkPlan = {  
11.     SparkPlan.currentContext.set(self)  
12. //SparkPlanner把LogicalPlan转化为SparkPlan  
13. //1.4.1选取的是第一个strategies DataSourceStrategy  
14.     planner.plan(optimizedPlan).next()  
15.   }  
16.   lazy val executedPlan: SparkPlan = prepareForExecution.execute(sparkPlan)  
17.   lazy val toRdd: RDD[Row] = {  
18.     toString  
19.     executedPlan.execute()  
20.   }  
21. protected def stringOrError[A](f: => A): String =  
22. try f.toString catch { case e: Throwable => e.toString }  
23.   def simpleString: String =  
24. """== Physical Plan ==  
25.        |${stringOrError(executedPlan)}  
26. """.stripMargin.trim  
27. //TODO:如何打印  
28.   override def toString: String = {  
29.     def output =  
30. "${o.name}: ${o.dataType.simpleString}").mkString(", ")  
31. // TODO previously will output RDD details by run (${stringOrError(toRdd.toDebugString)})  
32. // however, the `toRdd` will cause the real execution, which is not what we want.  
33. // We need to think about how to avoid the side effect.  
34. """== Parsed Logical Plan ==  
35.        |${stringOrError(logical)}  
36.        |== Analyzed Logical Plan ==  
37.        |${stringOrError(output)}  
38.        |${stringOrError(analyzed)}  
39.        |== Optimized Logical Plan ==  
40.        |${stringOrError(optimizedPlan)}  
41.        |== Physical Plan ==  
42.        |${stringOrError(executedPlan)}  
43.        |Code Generation: ${stringOrError(executedPlan.codegenEnabled)}  
44.        |== RDD ==  
45. """.stripMargin.trim  
46.   }  
47. }

这里唯一需要注意的是analyzed，optimizedPlan，sparkPlan，executedPlan都为懒变量，也就是说只有真正要用到的时时候才会去执行相应的代码逻辑，没有用到的时候是不会发生任何事情的。

1.4 LogicalPlan and SparkPlan

LogicalPlan和SparkPlan都继承自QueryPlan，QueryPlan为泛型类

1. abstract class QueryPlan[PlanType <: TreeNode[PlanType]] extends TreeNode[PlanType] {  
2. }  
3. abstract class LogicalPlan extends QueryPlan[LogicalPlan] with Logging {  
4. }  
5. abstract class SparkPlan extends QueryPlan[SparkPlan] with Logging with Serializable {  
6. }

以上都为抽象类，然后在此基础上又根据不同的类型衍生出不同的树节点

1. /**
2.  * A logical plan node with no children.叶子节点，没有子节点
3.  */  
4. abstract class LeafNode extends LogicalPlan with trees.LeafNode[LogicalPlan] {  
5.   self: Product =>  
6. }  
7. /**
8.  * A logical plan node with single child. 一元节点
9.  */  
10. abstract class UnaryNode extends LogicalPlan with trees.UnaryNode[LogicalPlan] {  
11.   self: Product =>  
12. }  
13. /**
14.  * A logical plan node with a left and right child 二元节点.
15.  */  
16. abstract class BinaryNode extends LogicalPlan with trees.BinaryNode[LogicalPlan] {  
17.   self: Product =>  
18. }

1. //叶子节点，没有子节点  
2. private[sql] trait LeafNode extends SparkPlan with trees.LeafNode[SparkPlan] {  
3.   self: Product =>  
4. }  
5. //一元节点  
6. private[sql] trait UnaryNode extends SparkPlan with trees.UnaryNode[SparkPlan] {  
7.   self: Product =>  
8. }  
9. //二元节点  
10. private[sql] trait BinaryNode extends SparkPlan with trees.BinaryNode[SparkPlan] {  
11.   self: Product =>  
12. }

其各自真正的具体类为：

1. abstract class LeafNode extends LogicalPlan with trees.LeafNode[LogicalPlan] {  
2.   self: Product =>  
3. }

Spark SQL发展历程 spark sql原理_sql_04

1. abstract class UnaryNode extends LogicalPlan with trees.UnaryNode[LogicalPlan] {  
2.   self: Product =>  
3. }

Spark SQL发展历程 spark sql原理_SQL_05

1. abstract class BinaryNode extends LogicalPlan with trees.BinaryNode[LogicalPlan] {  
2.   self: Product =>  
3. }

Spark SQL发展历程 spark sql原理_sql_06

1. private[sql] trait LeafNode extends SparkPlan with trees.LeafNode[SparkPlan] {  
2.   self: Product =>  
3. }

Spark SQL发展历程 spark sql原理_Spark SQL发展历程_07

1. private[sql] trait UnaryNode extends SparkPlan with trees.UnaryNode[SparkPlan] {  
2.   self: Product =>  
3. }

Spark SQL发展历程 spark sql原理_ide_08

1. private[sql] trait BinaryNode extends SparkPlan with trees.BinaryNode[SparkPlan] {  
2.   self: Product =>  
3. }

Spark SQL发展历程 spark sql原理_ide_09

可见Spark-Sql里面二叉树结构贯穿了整个解析过程。

二. Catalyst

所有的SQL操作最终都通过Catalyst翻译成spark程序代码

Spark SQL发展历程 spark sql原理_ide_10

三. SparkSQL整体架构（前端+后端）

Spark SQL发展历程 spark sql原理_SQL_11

thriftserver作为一个前端，它其实只是主要分为两大块：

1.维护与用户的JDBC连接

2.通过HiveContext的API提交用户的HQL查询

正因为当初对未来做了太多的憧憬，所以对现在的自己尤其失望。生命中曾经有过的所有灿烂，终究都需要用寂寞来偿还。

本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。

上一篇：hive入门 hive基础知识

下一篇：python将数组转换为列表 python数组转化为list

提问和评论都可以，用心的回复会被更多人看到评论

发布评论

相关文章

官方博客	全部文章	热门标签	班级博客
了解我们	网站地图	意见反馈

鸿蒙开发者社区	51CTO学堂
51CTO	软考资讯