大数据计算-SQL优化手段(CBO)-以Flink为例

背景

大数据计算中，SQL生成的执行计划第一轮会经过固定规则的优化，第二轮会根据原计划，生成多条结合成本的的执行计划，根据cost 进行排序，选出最优的执行计划。

理论知识

原始计划如左图，
有三种执行方案
方案1，scan表1，scan表2，然后hash ，再join
方案2，scan表1，scan表2，然后broadcast 表1 ，再join
方案2，scan表1，scan表2，然后broadcast 表2 ，再join

从成本(只看行数)来看，如果表aa_user 行数远小于bb_order ,那 方案2得出来的成本就是最优的。
下面是示意图

示例

aa_user 的表行数远小于bb_order

public static void main(String[] args) {EnvironmentSettings settings = EnvironmentSettings.inBatchMode();TableEnvironment tableEnvironment = TableEnvironment.create(settings);Schema schema = Schema.newBuilder().column("count", DataTypes.INT()).column("word", DataTypes.STRING()).build();Schema schema1 = Schema.newBuilder().column("id", DataTypes.INT()).column("name", DataTypes.STRING()).build();tableEnvironment.createTemporaryTable("aa_user", TableDescriptor.forConnector("filesystem").schema(schema).option("path","/Users/xx/IdeaProjects/flink-demo/data/order.csv").format("csv").build());tableEnvironment.createTemporaryTable("bb_order", TableDescriptor.forConnector("filesystem").schema(schema1).option("path","/Users/xx/IdeaProjects/flink-demo/data/user.csv").format("csv").build());//  tableEnvironment.executeSql("select * from aa_user").print();//tableEnvironment.executeSql("select * from aa_user inner join bb_order on `aa_user`.`count`=`bb_order`.`id`").print();String cost=    tableEnvironment.explainSql("select * from aa_user inner join bb_order on `aa_user`.`count`=`bb_order`.`id`", ExplainDetail.ESTIMATED_COST);System.out.println(cost);}

结果展示

== Abstract Syntax Tree ==
LogicalProject(count=[$0], word=[$1], id=[$2], name=[$3])
+- LogicalJoin(condition=[=($0, $2)], joinType=[inner]):- LogicalTableScan(table=[[default_catalog, default_database, aa_user]])+- LogicalTableScan(table=[[default_catalog, default_database, bb_order]])== Optimized Physical Plan ==
NestedLoopJoin(joinType=[InnerJoin], where=[=(count, id)], select=[count, word, id, name], build=[left]): rowcount = 87.6, cumulative cost = {673.6 rows, 1484.0 cpu, 9344.0 io, 32.0 network, 40.0 memory}
:- Exchange(distribution=[broadcast]): rowcount = 2.0, cumulative cost = {4.0 rows, 320.0 cpu, 32.0 io, 32.0 network, 0.0 memory}
:  +- TableSourceScan(table=[[default_catalog, default_database, aa_user]], fields=[count, word]): rowcount = 2.0, cumulative cost = {2.0 rows, 0.0 cpu, 32.0 io, 0.0 network, 0.0 memory}
+- TableSourceScan(table=[[default_catalog, default_database, bb_order]], fields=[id, name]): rowcount = 582.0, cumulative cost = {582.0 rows, 0.0 cpu, 9312.0 io, 0.0 network, 0.0 memory}== Optimized Execution Plan ==
MultipleInput(readOrder=[0,1], members=[\nNestedLoopJoin(joinType=[InnerJoin], where=[(count = id)], select=[count, word, id, name], build=[left])\n:- [#1] Exchange(distribution=[broadcast])\n+- [#2] TableSourceScan(table=[[default_catalog, default_database, bb_order]], fields=[id, name])\n])
:- Exchange(distribution=[broadcast])
:  +- TableSourceScan(table=[[default_catalog, default_database, aa_user]], fields=[count, word])
+- TableSourceScan(table=[[default_catalog, default_database, bb_order]], fields=[id, name])

结果解释

NestedLoopJoin：Flink 选择了嵌套循环连接（Nested Loop Join）作为执行 JOIN 的策略，使用 count = id 作为连接条件。
Exchange(distribution=[broadcast])：表示将 aa_user 表的数据广播分发，以减少数据移动的开销，rowcount = 2.0 表示预估的行数。
TableSourceScan：直接扫描表 aa_user 和 bb_order，并读取相应的字段。表 aa_user 预估有 2 行，表 bb_order 预估有 582 行