Spark Sql JDBC实现聚合、union、同数据源Join等下推

简单熟悉下Spark Sql 处理JDBC数据源数据

spark Sql处理JDBC数据源的代码比较简单，大家可以自行阅读官网使用demo。

个人觉得比较鸡肋的地方

单元测试如下：

说明： emp，dept是通过spark 读取mysql的同一个库的两张表

test("selectSubQuery"){    val sql =      """        |select *        |from   emp where id > (select min(a.id) from emp a join dept b on a.deptId =b.id)        |""".stripMargin    val frameResult = session.sql(sql)    frameResult.explain(true)    frameResult.show()  }

查看执行计划

== Parsed Logical Plan =='Project [*]+- 'Filter ('id > scalar-subquery#34 [])   :  +- 'Project [unresolvedalias('min('a.id), None)]   :     +- 'Join Inner, ('a.deptId = 'b.id)   :        :- 'SubqueryAlias a   :        :  +- 'UnresolvedRelation [emp], [], false   :        +- 'SubqueryAlias b   :           +- 'UnresolvedRelation [dept], [], false   +- 'UnresolvedRelation [emp], [], false== Analyzed Logical Plan ==ID: int, NAME: string, JOB: string, MGR: int, HIREDATE: date, SAL: decimal(20,2), COMM: decimal(20,2), DEPTID: intProject [ID#12, NAME#13, JOB#14, MGR#15, HIREDATE#16, SAL#17, COMM#18, DEPTID#19]+- Filter (id#12 > scalar-subquery#34 [])   :  +- Aggregate [min(id#12) AS min(id)#36]   :     +- Join Inner, (deptId#19 = id#6)   :        :- SubqueryAlias a   :        :  +- SubqueryAlias emp   :        :     +- Relation[ID#12,NAME#13,JOB#14,MGR#15,HIREDATE#16,SAL#17,COMM#18,DEPTID#19] JDBCRelation(emp) [numPartitions=1]   :        +- SubqueryAlias b   :           +- SubqueryAlias dept   :              +- Relation[ID#6,NAME#7,LOC#8] JDBCRelation(dept) [numPartitions=1]   +- SubqueryAlias emp      +- Relation[ID#12,NAME#13,JOB#14,MGR#15,HIREDATE#16,SAL#17,COMM#18,DEPTID#19] JDBCRelation(emp) [numPartitions=1]== Optimized Logical Plan ==Filter (isnotnull(id#12) AND (id#12 > scalar-subquery#34 [])):  +- Aggregate [min(id#12) AS min(id)#36]:     +- Project [ID#12]:        +- Join Inner, (deptId#19 = id#6):           :- Project [ID#12, DEPTID#19]:           :  +- Filter isnotnull(deptId#19):           :     +- Relation[ID#12,NAME#13,JOB#14,MGR#15,HIREDATE#16,SAL#17,COMM#18,DEPTID#19] JDBCRelation(emp) [numPartitions=1]:           +- Project [ID#6]:              +- Filter isnotnull(id#6):                 +- Relation[ID#6,NAME#7,LOC#8] JDBCRelation(dept) [numPartitions=1]+- Relation[ID#12,NAME#13,JOB#14,MGR#15,HIREDATE#16,SAL#17,COMM#18,DEPTID#19] JDBCRelation(emp) [numPartitions=1]== Physical Plan ==*(1) Filter (ID#12 > Subquery scalar-subquery#34, [id=#64]):  +- Subquery scalar-subquery#34, [id=#64]:     +- *(6) HashAggregate(keys=[], functions=[min(id#12)], output=[min(id)#36]):        +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [id=#60]:           +- *(5) HashAggregate(keys=[], functions=[partial_min(id#12)], output=[min#46]):              +- *(5) Project [ID#12]:                 +- *(5) SortMergeJoin [deptId#19], [id#6], Inner:                    :- *(2) Sort [deptId#19 ASC NULLS FIRST], false, 0:                    :  +- Exchange hashpartitioning(deptId#19, 200), ENSURE_REQUIREMENTS, [id=#45]:                    :     +- *(1) Scan JDBCRelation(emp) [numPartitions=1] [ID#12,DEPTID#19] PushedFilters: [*IsNotNull(DEPTID)], ReadSchema: struct
   
    :                    +- *(4) Sort [id#6 ASC NULLS FIRST], false, 0:                       +- Exchange hashpartitioning(id#6, 200), ENSURE_REQUIREMENTS, [id=#51]:                          +- *(3) Scan JDBCRelation(dept) [numPartitions=1] [ID#6] PushedFilters: [*IsNotNull(ID)], ReadSchema: struct
    
     +- *(1) Scan JDBCRelation(emp) [numPartitions=1] [ID#12,NAME#13,JOB#14,MGR#15,HIREDATE#16,SAL#17,COMM#18,DEPTID#19] PushedFilters: [*IsNotNull(ID)], ReadSchema: struct

执行结果如下：

缺点分析：

上图可以看出：

1.spark 总共进行了三次JDBC 连接进行查询请求。

2.聚合，join没有进行下沉，增加了内存计算的复杂度。

再次说明下，单元测试中的查询请求，比较简单，

就是查询同一个库的两张表格，会进行三次访问数据库，并且没有做到聚合和join的下沉（spark 目前支持了谓词下推，列裁剪，常量累加等优化），频繁的查库和数据的网络Io比较消耗时间。

优化操作

这里的优化主要使用的是Spark 提供的扩展接口Extensions，优化spark 的RBO和CBO规则。这里的原理和细节比较复杂，下篇文章描述。

优化后的执行计划

== Parsed Logical Plan =='Project [*]+- 'Filter ('id > scalar-subquery#34 [])   :  +- 'Project [unresolvedalias('min('a.id), None)]   :     +- 'Join Inner, ('a.deptId = 'b.id)   :        :- 'SubqueryAlias a   :        :  +- 'UnresolvedRelation [emp], [], false   :        +- 'SubqueryAlias b   :           +- 'UnresolvedRelation [dept], [], false   +- 'UnresolvedRelation [emp], [], false== Analyzed Logical Plan ==ID: int, NAME: string, JOB: string, MGR: int, HIREDATE: date, SAL: decimal(20,2), COMM: decimal(20,2), DEPTID: intProject [ID#12, NAME#13, JOB#14, MGR#15, HIREDATE#16, SAL#17, COMM#18, DEPTID#19]+- Filter (id#12 > scalar-subquery#34 [])   :  +- Aggregate [min(id#12) AS min(id)#36]   :     +- Join Inner, (deptId#19 = id#6)   :        :- SubqueryAlias a   :        :  +- SubqueryAlias emp   :        :     +- RelationV2[ID#12, NAME#13, JOB#14, MGR#15, HIREDATE#16, SAL#17, COMM#18, DEPTID#19] nameString   :        +- SubqueryAlias b   :           +- SubqueryAlias dept   :              +- RelationV2[ID#6, NAME#7, LOC#8] nameString   +- SubqueryAlias emp      +- RelationV2[ID#12, NAME#13, JOB#14, MGR#15, HIREDATE#16, SAL#17, COMM#18, DEPTID#19] nameString== Optimized Logical Plan ==RelationV2[ID#47, NAME#48, JOB#49, MGR#50, HIREDATE#51, SAL#52, COMM#53, DEPTID#54] nameString== Physical Plan ==*(1) Scan org.apache.spark.sql.execution.datasources.v2.jdbc.JDBCScan$$anon$1@237c8718 [ID#47,NAME#48,JOB#49,MGR#50,HIREDATE#51,SAL#52,COMM#53,DEPTID#54] PushedFilters: [], ReadSchema: struct

在这里插入图片描述