flink sql

时间  2020-03-28
标签 flink sql 栏目 SQL 繁體版
原文   原文链接

StreamTableEnvironment

该类包含sql解析、验证、优化、执行等各环节须要的元数据管理器CatalogManager,模块管理器(模块包含函数集、类型集、规则集)moduleManager，用户自定义函数管理器FunctionCatalog,线程池、sql解析器Planner。node

StreamTableEnvironmentImpl.create(executionEnvironment, settings, new TableConfig)

  def create(
      executionEnvironment: StreamExecutionEnvironment,
      settings: EnvironmentSettings,
      tableConfig: TableConfig)
    : StreamTableEnvironmentImpl = {

    val catalogManager = new CatalogManager(
      settings.getBuiltInCatalogName,
      new GenericInMemoryCatalog(settings.getBuiltInCatalogName, settings.getBuiltInDatabaseName))

    val moduleManager = new ModuleManager
    val functionCatalog = new FunctionCatalog(catalogManager, moduleManager)

    val executorProperties = settings.toExecutorProperties
    val executor = lookupExecutor(executorProperties, executionEnvironment)

    val plannerProperties = settings.toPlannerProperties
    val planner = ComponentFactoryService.find(classOf[PlannerFactory], plannerProperties)
      .create(
        plannerProperties,
        executor,
        tableConfig,
        functionCatalog,
        catalogManager)

    new StreamTableEnvironmentImpl(
      catalogManager,
      moduleManager,
      functionCatalog,
      tableConfig,
      executionEnvironment,
      planner,
      executor,
      settings.isStreamingMode
    )
  }

DataType

定义了逻辑类型，而且对其底层实际物理类型进行暗示。sql

LogicalType

逻辑类型有点相似标准SQL的数据类型,其子类作了具体的约束。express

TableSchema

表结构定义，包含各字段名称和各字段类型c#

DataStream -> Table

override def fromDataStream[T](dataStream: DataStream[T], fields: Expression*): Table = {
    val queryOperation = asQueryOperation(dataStream, Some(fields.toList.asJava))
    createTable(queryOperation)
  }

ScalaDataStreamQueryOperation

private final DataStream<E> dataStream;
	private final int[] fieldIndices;
	private final TableSchema tableSchema;

Table

Table类是sql api的核心组件，定义了转换数据的方法如filter、groupBy、join等。使用TableEnvironment类能够把Table转换成DataStream或者DataSet。api

private TableImpl(
			TableEnvironment tableEnvironment,
			QueryOperation operationTree,
			OperationTreeBuilder operationTreeBuilder,
			LookupCallResolver lookupResolver) {
		this.tableEnvironment = tableEnvironment;
		this.operationTree = operationTree;
		this.operationTreeBuilder = operationTreeBuilder;
		this.lookupResolver = lookupResolver;
	}

注册表信息

private void createTemporaryView(UnresolvedIdentifier identifier, Table view) {
		if (((TableImpl) view).getTableEnvironment() != this) {
			throw new TableException(
				"Only table API objects that belong to this TableEnvironment can be registered.");
		}

		CatalogBaseTable tableTable = new QueryOperationCatalogView(view.getQueryOperation());

		ObjectIdentifier tableIdentifier = catalogManager.qualifyIdentifier(identifier);
		catalogManager.createTemporaryTable(tableTable, tableIdentifier, false);
	}

Expression

Expression表明字面量、函数调用或者field引用。ide

ExpressionVisitor

转换Expression的visitor函数

IndexedExprToFieldInfo

ExpressionVisitor的子类把Expression解析成FieldInfo优化

@Override
		public FieldInfo visit(UnresolvedReferenceExpression unresolvedReference) {
			String fieldName = unresolvedReference.getName();
			return new FieldInfo(fieldName, index, fromLegacyInfoToDataType(getTypeAt(unresolvedReference)));
		}

应用举例,把Expression转换成FieldInfo:ui

private static List<FieldInfo> extractFieldInfosFromTupleType(TupleTypeInfoBase<?> inputType, Expression[] exprs) {
		boolean isRefByPos = isReferenceByPosition(inputType, exprs);

		if (isRefByPos) {
			return IntStream.range(0, exprs.length)
				.mapToObj(idx -> exprs[idx].accept(new IndexedExprToFieldInfo(inputType, idx)))
				.collect(Collectors.toList());
		} else {
			return extractFieldInfosByNameReference(inputType, exprs);
		}
	}

FieldInfo

private final String fieldName;
		private final int index;
		private final DataType type;

Row & RowTypeInfo

表明一行数据，能够包含任意数量的列，而且各列可能包含不一样的数据类型.Row不是强类型的因此须要配合RowTypeInfo类获取各列具体的类型.this

Row:

/** The array to store actual values. */
	private final Object[] fields;

RowTypeInfo:

protected final String[] fieldNames;
	protected final TypeInformation<?>[] types;

Table -> DataStream

override def toAppendStream[T: TypeInformation](table: Table): DataStream[T] = {
    val returnType = createTypeInformation[T]

    val modifyOperation = new OutputConversionModifyOperation(
      table.getQueryOperation,
      TypeConversions.fromLegacyInfoToDataType(returnType),
      OutputConversionModifyOperation.UpdateMode.APPEND)
    toDataStream[T](table, modifyOperation)
  }

Operation

Parser.parse(sql)的返回结果。

  • ModifyOperation (DML)
  • QueryOperation (DQL)
  • CreateOperation & DropOperation (DDL)

FlinkStreamRuleSets

定义了sql解析优化规则集合,包含把calcite节点转换成flink节点的规则，好比FlinkLogicalTableSourceScan,把flink逻辑节点转换成物理执行节点的规则,好比StreamExecTableSourceScanRule，条件过滤下推的规则PushFilterIntoTableSourceScanRule等.

ConverterRule

/** Converts a relational expression to the target trait(s) of this rule.
   	*
   	* <p>Returns null if conversion is not possible. */
  	public abstract RelNode convert(RelNode rel);
  
    public void onMatch(RelOptRuleCall call) {
    RelNode rel = call.rel(0);
    if (rel.getTraitSet().contains(inTrait)) {
      final RelNode converted = convert(rel);
      if (converted != null) {
        call.transformTo(converted);
      }
    }
  }
  
  class FlinkLogicalTableSourceScanConverter
  extends ConverterRule(
    classOf[LogicalTableScan],
    Convention.NONE,
    FlinkConventions.LOGICAL,
    "FlinkLogicalTableSourceScanConverter") {

  override def matches(call: RelOptRuleCall): Boolean = {
    val scan: TableScan = call.rel(0)
    isTableSourceScan(scan)
  }

  def convert(rel: RelNode): RelNode = {
    val scan = rel.asInstanceOf[TableScan]
    val table = scan.getTable.asInstanceOf[FlinkRelOptTable]
    FlinkLogicalTableSourceScan.create(rel.getCluster, table)
  }
}

FlinkLogicalRel

flink RelNode基类不只包含了RelNode自己可表达的关系依赖逻辑，并且包含了各关系依赖的Flink体系中的额外信息。好比FlinkLogicalTableSourceScan包含了TableSource信息。

FlinkPhysicalRel

物理关系节点基类,其子类同时也会实现ExecNode接口，用于把物理节点转换成Transformation

ExecNode

/**
    * Internal method, translates this node into a Flink operator.
    *
    * @param planner The [[Planner]] of the translated Table.
    */
  protected def translateToPlanInternal(planner: E): Transformation[T]
  
  
  def translateToPlan(planner: E): Transformation[T] = {
    if (transformation == null) {
      transformation = translateToPlanInternal(planner)
    }
    transformation
  }

调用时序图

代码生成gencode

ExecNode转换成Transformation的过程当中部分逻辑会采用动态生成代码的方式实现。动态生成的代码保存在GeneratedClass子类的实例中，会分发到各个TM节点而后由Janino库编译执行。好比聚合查询生成聚合处理函数NamespaceTableAggsHandleFunction的子类。

GeneratedClass

public T newInstance(ClassLoader classLoader, Object... args) {
		try {
			return (T) compile(classLoader).getConstructors()[0].newInstance(args);
		} catch (Exception e) {
			throw new RuntimeException(
					"Could not instantiate generated class '" + className + "'", e);
		}
	}

	/**
	 * Compiles the generated code, the compiled class will be cached in the {@link GeneratedClass}.
	 */
	public Class<T> compile(ClassLoader classLoader) {
		if (compiledClass == null) {
			// cache the compiled class
			compiledClass = CompileUtils.compile(classLoader, className, code);
		}
		return compiledClass;
	}

示例

val sql =
      """
        |SELECT
        |  `string`,
        |  HOP_START(rowtime, INTERVAL '0.004' SECOND, INTERVAL '0.005' SECOND),
        |  HOP_ROWTIME(rowtime, INTERVAL '0.004' SECOND, INTERVAL '0.005' SECOND),
        |  COUNT(1),
        |  SUM(1),
        |  COUNT(`int`),
        |  COUNT(DISTINCT `float`),
        |  concat_distinct_agg(name)
        |FROM T1
        |GROUP BY `string`, HOP(rowtime, INTERVAL '0.004' SECOND, INTERVAL '0.005' SECOND)
      """.stripMargin
LogicalProject#3
	LogicalAggregate#2
		LogicalProject#1
			LogicalTableScan#0
rel#271:StreamExecSink.STREAM_PHYSICAL.any.None: 0.false.Acc(input=StreamExecCalc#269,name=DataStreamTableSink,fields=string, EXPR$1, EXPR$2, EXPR$3, EXPR$4, EXPR$5, EXPR$6, EXPR$7)
	rel#269:StreamExecCalc.STREAM_PHYSICAL.any.None: 0.false.Acc(input=StreamExecGroupWindowAggregate#267,select=string, w$start AS EXPR$1, w$rowtime AS EXPR$2, EXPR$3, EXPR$4, EXPR$5, EXPR$6, EXPR$7)
		rel#267:StreamExecGroupWindowAggregate.STREAM_PHYSICAL.any.None: 0.false.Acc(input=StreamExecExchange#265,groupBy=string,window=SlidingGroupWindow('w$, rowtime, 5, 4),properties=w$start, w$end, w$rowtime, w$proctime,select=string, COUNT(*) AS EXPR$3, $SUM0($f2) AS EXPR$4, COUNT(int) AS EXPR$5, COUNT(DISTINCT float) AS EXPR$6, concat_distinct_agg(name) AS EXPR$7, start('w$) AS w$start, end('w$) AS w$end, rowtime('w$) AS w$rowtime, proctime('w$) AS w$proctime)
			rel#265:StreamExecExchange.STREAM_PHYSICAL.hash[0]true.None: -1.true.Acc(input=StreamExecCalc#263,distribution=hash[string])
				rel#263:StreamExecCalc.STREAM_PHYSICAL.any.None: -1.true.Acc(input=StreamExecDataStreamScan#257,select=string, rowtime, 1 AS $f2, int, float, name)
					rel#257:StreamExecDataStreamScan.STREAM_PHYSICAL.any.None: -1.true.Acc(table=[Unregistered_DataStream_2],fields=rowtime, int, double, float, bigdec, string, name)

代码生成:

StreamExecGroupWindowAggregateBase->translateToPlanInternal
	StreamExecGroupWindowAggregateBase ->createAggsHandler
		AggsHandlerCodeGenerator->generateNamespaceAggsHandler
			new OneInputTransformation
			
任务提交中会把 OneInputTransformation -> OneInputStreamTask			
				
Task->run
	Task->doRun
		StreamTask->invoke
			StreamTask->openAllOperators
				AggregateWindowOperator->open
					WindowOperator->compileGeneratedCode
						GeneratedNamespaceAggsHandleFunction->newInstance
							SimpleCompiler->cook
相关文章
相关标签/搜索
每日一句
    每一个你不满意的现在，都有一个你没有努力的曾经。
本站公众号
   欢迎关注本站公众号,获取更多信息
联系我们 最近搜索 最新文章 沪ICP备13005482号-10 MyBatis教程 SQL 教程 MySQL教程 Java 教程 Thymeleaf 教程 Hibernate教程 Spring教程 Redis教程