Spark SQL 的 Catalyst ,这部分真的很有意思,值得去仔细研究一番,今天先来说说Spark的一些扩展机制吧,上一次写Spark,对其SQL的解析进行了一定的魔改,今天我们按套路来,使用砖厂为我们提供的机制,来扩展Spark...
首先我们先来了解一下 Spark SQL 的整体执行流程,输入的查询先被解析成未关联元数据的逻辑计划,然后根据元数据和解析规则,生成逻辑计划,再经过优化规则,形成优化过的逻辑计划(RBO),将逻辑计划转换成物理计划在经过代价模型(CBO),输出真正的物理执行计划。
我们今天举三个扩展的例子,来进行说明。
扩展解析器
这个例子,我们扩展解析引擎,我们对输入的SQL,禁止泛查询即不许使用select *来做查询,以下是解析的代。
- package wang.datahub.parser
- import org.apache.spark.sql.catalyst.analysis.UnresolvedStar
- import org.apache.spark.sql.catalyst.expressions.Expression
- import org.apache.spark.sql.catalyst.parser.ParserInterface
- import org.apache.spark.sql.catalyst.plans.logical.{LogicalPlan, Project}
- import org.apache.spark.sql.catalyst.{FunctionIdentifier, TableIdentifier}
- import org.apache.spark.sql.types.{DataType, StructType}
- class MyParser(parser: ParserInterface) extends ParserInterface {
- /**
- * Parse a string to a [[LogicalPlan]].
- */
- override def parsePlan(sqlText: String): LogicalPlan = {
- val logicalPlan = parser.parsePlan(sqlText)
- logicalPlan transform {
- case project @ Project(projectList, _) =>
- projectList.foreach {
- name =>
- if (name.isInstanceOf[UnresolvedStar]) {
- throw new RuntimeException("You must specify your project column set," +
- " * is not allowed.")
- }
- }
- project
- }
- logicalPlan
- }
- /**
- * Parse a string to an [[Expression]].
- */
- override def parseExpression(sqlText: String): Expression = parser.parseExpression(sqlText)
- /**
- * Parse a string to a [[TableIdentifier]].
- */
- override def parseTableIdentifier(sqlText: String): TableIdentifier =
- parser.parseTableIdentifier(sqlText)
- /**
- * Parse a string to a [[FunctionIdentifier]].
- */
- override def parseFunctionIdentifier(sqlText: String): FunctionIdentifier =
- parser.parseFunctionIdentifier(sqlText)
- /**
- * Parse a string to a [[StructType]]. The passed SQL string should be a comma separated
- * list of field definitions which will preserve the correct Hive metadata.
- */
- override def parseTableSchema(sqlText: String): StructType =
- parser.parseTableSchema(sqlText)
- /**
- * Parse a string to a [[DataType]].
- */
- override def parseDataType(sqlText: String): DataType = parser.parseDataType(sqlText)
- }
接下来,我们测试一下
- package wang.datahub.parser
- import org.apache.spark.sql.{SparkSession, SparkSessionExtensions}
- import org.apache.spark.sql.catalyst.parser.ParserInterface
- object MyParserApp {
- def main(args: Array[String]): Unit = {
- System.setProperty("hadoop.home.dir","E:\\devlop\\envs\\hadoop-common-2.2.0-bin-master");
- type ParserBuilder = (SparkSession, ParserInterface) => ParserInterface
- type ExtensionsBuilder = SparkSessionExtensions => Unit
- val parserBuilder: ParserBuilder = (_, parser) => new MyParser(parser)
- val extBuilder: ExtensionsBuilder = { e => e.injectParser(parserBuilder)}
- val spark = SparkSession
- .builder()
- .appName("Spark SQL basic example")
- .config("spark.master", "local[*]")
- .withExtensions(extBuilder)
- .getOrCreate()
- spark.sparkContext.setLogLevel("ERROR")
- import spark.implicits._
- val df = Seq(
- ( "First Value",1, java.sql.Date.valueOf("2010-01-01")),
- ( "First Value",4, java.sql.Date.valueOf("2010-01-01")),
- ("Second Value",2, java.sql.Date.valueOf("2010-02-01")),
- ("Second Value",9, java.sql.Date.valueOf("2010-02-01"))
- ).toDF("name", "score", "date_column")
- df.createTempView("p")
- // val df = spark.read.json("examples/src/main/resources/people.json")
- // df.toDF().write.saveAsTable("person")
- //,javg(score)
- // custom parser
- // spark.sql("select * from p ").show
- spark.sql("select * from p").show()
- }
- }
下面是执行结果,符合我们的预期。
扩展优化器
接下来,我们来扩展优化器,砖厂提供了很多默认的RBO,这里可以方便的构建我们自己的优化规则,本例中我们构建一套比较奇怪的规则,而且是完全不等价的,这里只是为了说明。
针对字段+0的操作,规则如下:
- 如果0出现在+左边,则直接将字段变成右表达式,即 0+nr 等效为 nr
- 如果0出现在+右边,则将0变成3,即 nr+0 变成 nr+3
- 如果没出现0,则表达式不变
下面是代码:
- package wang.datahub.optimizer
- import org.apache.spark.sql.SparkSession
- import org.apache.spark.sql.catalyst.expressions.{Add, Expression, Literal}
- import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
- import org.apache.spark.sql.catalyst.rules.Rule
- object MyOptimizer extends Rule[LogicalPlan] {
- def apply(logicalPlan: LogicalPlan): LogicalPlan = {
- logicalPlan.transformAllExpressions {
- case Add(left, right) => {
- println("this this my add optimizer")
- if (isStaticAdd(left)) {
- right
- } else if (isStaticAdd(right)) {
- Add(left, Literal(3L))
- } else {
- Add(left, right)
- }
- }
- }
- }
- private def isStaticAdd(expression: Expression): Boolean = {
- expression.isInstanceOf[Literal] && expression.asInstanceOf[Literal].toString == "0"
- }
- def main(args: Array[String]): Unit = {
- System.setProperty("hadoop.home.dir","E:\\devlop\\envs\\hadoop-common-2.2.0-bin-master");
- val testSparkSession: SparkSession = SparkSession.builder().appName("Extra optimization rules")
- .master("local[*]")
- .withExtensions(extensions => {
- extensions.injectOptimizerRule(session => MyOptimizer)
- })
- .getOrCreate()
- testSparkSession.sparkContext.setLogLevel("ERROR")
- import testSparkSession.implicits._
- testSparkSession.experimental.extraOptimizations = Seq()
- Seq(-1, -2, -3).toDF("nr").write.mode("overwrite").json("./test_nrs")
- // val optimizedResult = testSparkSession.read.json("./test_nrs").selectExpr("nr + 0")
- testSparkSession.read.json("./test_nrs").createTempView("p")
- var sql = "select nr+0 from p";
- var t = testSparkSession.sql(sql)
- println(t.queryExecution.optimizedPlan)
- println(sql)
- t.show()
- sql = "select 0+nr from p";
- var u = testSparkSession.sql(sql)
- println(u.queryExecution.optimizedPlan)
- println(sql)
- u.show()
- sql = "select nr+8 from p";
- var v = testSparkSession.sql(sql)
- println(v.queryExecution.optimizedPlan)
- println(sql)
- v.show()
- // println(optimizedResult.queryExecution.optimizedPlan.toString() )
- // optimizedResult.collect().map(row => row.getAs[Long]("(nr + 0)"))
- Thread.sleep(1000000)
- }
- }
执行如下
- this this my add optimizer
- this this my add optimizer
- this this my add optimizer
- Project [(nr#12L + 3) AS (nr + CAST(0 AS BIGINT))#14L]
- +- Relation[nr#12L] json
- select nr+0 from p
- this this my add optimizer
- this this my add optimizer
- this this my add optimizer
- +------------------------+
- |(nr + CAST(0 AS BIGINT))|
- +------------------------+
- | 2|
- | 1|
- | 0|
- +------------------------+
- this this my add optimizer
- Project [nr#12L AS (CAST(0 AS BIGINT) + nr)#21L]
- +- Relation[nr#12L] json
- select 0+nr from p
- this this my add optimizer
- +------------------------+
- |(CAST(0 AS BIGINT) + nr)|
- +------------------------+
- | -1|
- | -2|
- | -3|
- +------------------------+
- this this my add optimizer
- this this my add optimizer
- this this my add optimizer
- Project [(nr#12L + 8) AS (nr + CAST(8 AS BIGINT))#28L]
- +- Relation[nr#12L] json
- select nr+8 from p
- this this my add optimizer
- this this my add optimizer
- this this my add optimizer
- +------------------------+
- |(nr + CAST(8 AS BIGINT))|
- +------------------------+
- | 7|
- | 6|
- | 5|
- +------------------------+
扩展策略
SparkStrategies包含了一系列特定的Strategies,这些Strategies是继承自QueryPlanner中定义的Strategy,它定义接受一个Logical Plan,生成一系列的Physical Plan
通过Strategies把逻辑计划转换成可以具体执行的物理计划,代码如下
- package wang.datahub.strategy
- import org.apache.spark.sql.{SparkSession, Strategy}
- import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
- import org.apache.spark.sql.execution.SparkPlan
- object MyStrategy extends Strategy {
- def apply(plan: LogicalPlan): Seq[SparkPlan] = {
- println("Hello world!")
- Nil
- }
- def main(args: Array[String]): Unit = {
- System.setProperty("hadoop.home.dir","E:\\devlop\\envs\\hadoop-common-2.2.0-bin-master");
- val spark = SparkSession.builder().master("local").getOrCreate()
- spark.experimental.extraStrategies = Seq(MyStrategy)
- val q = spark.catalog.listTables.filter(t => t.name == "six")
- q.explain(true)
- spark.stop()
- }
- }
执行效果
好了,扩展部分就先介绍到这,接下来我计划可能会简单说说RBO和CBO,结合之前做过的一个小功能,一条SQL的查询时间预估。
本文转载自微信公众号「麒思妙想」,可以通过以下二维码关注。转载本文请联系麒思妙想公众号。