Flink源码阅读之Flinksql执行流程

最近工作中需要开发一些新的Flink sql 的connector，所以先开始研究研究Flink sql的执行流程。

基本结构

Planner接口
负责sql解析、转换成Transformation
Executor接口
负责将planner转换的Transformation生成streamGraph并执行

public interface Planner {

    /**
     * Retrieves a {@link Parser} that provides methods for parsing a SQL string.
     *
     * @return initialized {@link Parser}
     */
    Parser getParser();

    /**
     * Converts a relational tree of {@link ModifyOperation}s into a set of runnable
     * {@link Transformation}s.
     *
     * <p>This method accepts a list of {@link ModifyOperation}s to allow reusing common
     * subtrees of multiple relational queries. Each query's top node should be a {@link ModifyOperation}
     * in order to pass the expected properties of the output {@link Transformation} such as
     * output mode (append, retract, upsert) or the expected output type.
     *
     * @param modifyOperations list of relational operations to plan, optimize and convert in a
     * single run.
     * @return list of corresponding {@link Transformation}s.
     */
    List<Transformation<?>> translate(List<ModifyOperation> modifyOperations);

    /**
     * Returns the AST of the specified Table API and SQL queries and the execution plan
     * to compute the result of the given collection of {@link QueryOperation}s.
     *
     * @param operations The collection of relational queries for which the AST
     * and execution plan will be returned.
     * @param extended if the plan should contain additional properties such as
     * e.g. estimated cost, traits
     */
    String explain(List<Operation> operations, boolean extended);

    /**
     * Returns completion hints for the given statement at the given cursor position.
     * The completion happens case insensitively.
     *
     * @param statement Partial or slightly incorrect SQL statement
     * @param position cursor position
     * @return completion hints that fit at the current cursor position
     */
    String[] getCompletionHints(String statement, int position);
}

Sql解析

Parser接口
负责sql解析
有两个实现一个是old planner，另一个是blink planner
flink对sql的解析依赖于calcite

具体实现

@Override
    public List<Operation> parse(String statement) {
        CalciteParser parser = calciteParserSupplier.get();
        FlinkPlannerImpl planner = validatorSupplier.get();
        // parse the sql query
        //依赖calcite将sql语句解析为sqlNode
        SqlNode parsed = parser.parse(statement);

        //将sqlnode转换为Operation
        Operation operation = SqlToOperationConverter.convert(planner, catalogManager, parsed)
            .orElseThrow(() -> new TableException("Unsupported query: " + statement));
        return Collections.singletonList(operation);
    }
    
    /**
     * This is the main entrance for executing all kinds of DDL/DML {@code SqlNode}s, different
     * SqlNode will have it's implementation in the #convert(type) method whose 'type' argument
     * is subclass of {@code SqlNode}.
     *
     * @param flinkPlanner FlinkPlannerImpl to convertCreateTable sql node to rel node
     * @param catalogManager CatalogManager to resolve full path for operations
     * @param sqlNode SqlNode to execute on
     */
    public static Optional<Operation> convert(
            FlinkPlannerImpl flinkPlanner,
            CatalogManager catalogManager,
            SqlNode sqlNode) {
        // validate the query
        // 校验sql的合法性
        final SqlNode validated = flinkPlanner.validate(sqlNode);
        SqlToOperationConverter converter = new SqlToOperationConverter(flinkPlanner, catalogManager);
        //对不同的ddl/dml进行转换
        if (validated instanceof SqlCreateTable) {
            return Optional.of(converter.convertCreateTable((SqlCreateTable) validated));
        } else if (validated instanceof SqlDropTable) {
            return Optional.of(converter.convertDropTable((SqlDropTable) validated));
        } else if (validated instanceof SqlAlterTable) {
            return Optional.of(converter.convertAlterTable((SqlAlterTable) validated));
        } else if (validated instanceof SqlCreateFunction) {
            return Optional.of(converter.convertCreateFunction((SqlCreateFunction) validated));
        } else if (validated instanceof SqlAlterFunction) {
            return Optional.of(converter.convertAlterFunction((SqlAlterFunction) validated));
        } else if (validated instanceof SqlDropFunction) {
            return Optional.of(converter.convertDropFunction((SqlDropFunction) validated));
        } else if (validated instanceof RichSqlInsert) {
            return Optional.of(converter.convertSqlInsert((RichSqlInsert) validated));
        } else if (validated instanceof SqlUseCatalog) {
            return Optional.of(converter.convertUseCatalog((SqlUseCatalog) validated));
        } else if (validated instanceof SqlUseDatabase) {
            return Optional.of(converter.convertUseDatabase((SqlUseDatabase) validated));
        } else if (validated instanceof SqlCreateDatabase) {
            return Optional.of(converter.convertCreateDatabase((SqlCreateDatabase) validated));
        } else if (validated instanceof SqlDropDatabase) {
            return Optional.of(converter.convertDropDatabase((SqlDropDatabase) validated));
        } else if (validated instanceof SqlAlterDatabase) {
            return Optional.of(converter.convertAlterDatabase((SqlAlterDatabase) validated));
        } else if (validated instanceof SqlCreateCatalog) {
            return Optional.of(converter.convertCreateCatalog((SqlCreateCatalog) validated));
        } else if (validated.getKind().belongsTo(SqlKind.QUERY)) {
            return Optional.of(converter.convertSqlQuery(validated));
        } else {
            return Optional.empty();
        }
    }

举个栗子：

CREATE TABLE user_behavior (
                    user_id BIGINT,
                    item_id BIGINT,
                    category_id BIGINT,
                    behavior STRING,
                    ts TIMESTAMP(3),
                    proctime as PROCTIME(),   -- 通过计算列产生一个处理时间列
                    WATERMARK FOR ts as ts - INTERVAL '5' SECOND  -- 在ts上定义watermark，ts成为事件时间列
                ) WITH (
                    'connector.type' = 'kafka',  -- 使用 kafka connector
                    'connector.version' = 'universal',  -- kafka 版本，universal 支持 0.11 以上的版本
                    'connector.topic' = 'user_behavior',  -- kafka topic
                    'connector.startup-mode' = 'earliest-offset',  -- 从起始 offset 开始读取
                    --'connector.properties.group.id' = '',
                    'connector.properties.zookeeper.connect' = 'localhost:2181',  -- zookeeper 地址
                    'connector.properties.bootstrap.servers' = 'localhost:9092',  -- kafka broker 地址
                    'format.type' = 'json'  -- 数据源格式为 json
                )

上面的sql经过SqlNode parsed = parser.parse(statement);解析之后如下图：

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-GnFJR75e-1593768699547)(/Users/admin/Library/Containers/com.tencent.WeWorkMac/Data/Library/Application Support/WXWork/Temp/ScreenCapture/企业微信截图_13cee561-c9a0-4318-acd6-bd48a418549d.png)]

对应的表名、列名、with属性参数、主键、唯一键、分区键、水印、表注释、表操作（create table、alter table、drop table。。。）都放到SqlNode对象的对应属性中，SqlNode是一个树形结构也就是AST。
不同的sql经过convert处理后返回不同的Operation，最后会根据不同的Operation有不同的处理行为。
Operation主要分为Create/Alter/Drop/Modify/Query/Use 几大类，每个下面又细分如CreateOperation

在这里插入图片描述

看下createTable的处理，返回的就是CreateTableOperation

    /**
     * Convert the {@link SqlCreateTable} node.
     */
    private Operation convertCreateTable(SqlCreateTable sqlCreateTable) {
        // primary key and unique keys are not supported
        if ((sqlCreateTable.getPrimaryKeyList().size() > 0)
            || (sqlCreateTable.getUniqueKeysList().size() > 0)) {
            throw new SqlConversionException("Primary key and unique key are not supported yet.");
        }

        // set with properties
    //将with参数放到一个map中
        Map<String, String> properties = new HashMap<>();
        sqlCreateTable.getPropertyList().getList().forEach(p ->
            properties.put(((SqlTableOption) p).getKeyString(), ((SqlTableOption) p).getValueString()));
        //schema
        TableSchema tableSchema = createTableSchema(sqlCreateTable);
        String tableComment = sqlCreateTable.getComment().map(comment ->
            comment.getNlsString().getValue()).orElse(null);
        // set partition key
        List<String> partitionKeys = sqlCreateTable.getPartitionKeyList()
            .getList()
            .stream()
            .map(p -> ((SqlIdentifier) p).getSimple())
            .collect(Collectors.toList());

        CatalogTable catalogTable = new CatalogTableImpl(tableSchema,
            partitionKeys,
            properties,
            tableComment);

        UnresolvedIdentifier unresolvedIdentifier = UnresolvedIdentifier.of(sqlCreateTable.fullTableName());
        ObjectIdentifier identifier = catalogManager.qualifyIdentifier(unresolvedIdentifier);

        return new CreateTableOperation(
            identifier,
            catalogTable,
            sqlCreateTable.isIfNotExists());
    }

else if (operation instanceof CreateTableOperation) {
            CreateTableOperation createTableOperation = (CreateTableOperation) operation;
            catalogManager.createTable(
                createTableOperation.getCatalogTable(),
                createTableOperation.getTableIdentifier(),
                createTableOperation.isIgnoreIfExists());
        }

对表的DDL操作（比如对table、function、database的操作）入口为tableEnv.sqlUpdate(sql),会调用到catalogManager对表的DDL进行维护，如果是query语句入口tableEnv.sqlQuery(sql),则会走

else if (validated.getKind().belongsTo(SqlKind.QUERY)) {
            return Optional.of(converter.convertSqlQuery(validated));
        }

private PlannerQueryOperation toQueryOperation(FlinkPlannerImpl planner, SqlNode validated) {
        // transform to a relational tree
        RelRoot relational = planner.rel(validated);
        return new PlannerQueryOperation(relational.project());
    }

将SqlNode转换为RelNode，

主要包含3个步骤：

推断table类型
推断计算列

推断watermark分配

val tableSourceTable = new TableSourceTable[T](
    relOptSchema,
    schemaTable.getTableIdentifier,
    erasedRowType,
    statistic,
    tableSource,
    schemaTable.isStreamingMode,
    catalogTable)

  // 1. push table scan

  // Get row type of physical fields.
  val physicalFields = getRowType
    .getFieldList
    .filter(f => !columnExprs.contains(f.getName))
    .map(f => f.getIndex)
    .toArray
  // Copy this table with physical scan row type.
  val newRelTable = tableSourceTable.copy(tableSource, physicalFields)
  val scan = LogicalTableScan.create(cluster, newRelTable)
  val relBuilder = FlinkRelBuilder.of(cluster, getRelOptSchema)
  relBuilder.push(scan)

  val toRexFactory = cluster
      .getPlanner
      .getContext
      .unwrap(classOf[FlinkContext])
      .getSqlExprToRexConverterFactory

  // 2. push computed column project
  val fieldNames = erasedRowType.getFieldNames.asScala
  if (columnExprs.nonEmpty) {
    val fieldExprs = fieldNames
        .map { name =>
          if (columnExprs.contains(name)) {
            columnExprs(name)
          } else {
            s"`$name`"
          }
        }.toArray
    val rexNodes = toRexFactory.create(newRelTable.getRowType).convertToRexNodes(fieldExprs)
    relBuilder.projectNamed(rexNodes.toList, fieldNames, true)
  }

  // 3. push watermark assigner
  val watermarkSpec = catalogTable
    .getSchema
    // we only support single watermark currently
    .getWatermarkSpecs.asScala.headOption
  if (schemaTable.isStreamingMode && watermarkSpec.nonEmpty) {
    if (TableSourceValidation.hasRowtimeAttribute(tableSource)) {
      throw new TableException(
        "If watermark is specified in DDL, the underlying TableSource of connector" +
            " shouldn't return an non-empty list of RowtimeAttributeDescriptor" +
            " via DefinedRowtimeAttributes interface.")
    }
    val rowtime = watermarkSpec.get.getRowtimeAttribute
    if (rowtime.contains(".")) {
      throw new TableException(
        s"Nested field '$rowtime' as rowtime attribute is not supported right now.")
    }
    val rowtimeIndex = fieldNames.indexOf(rowtime)
    val watermarkRexNode = toRexFactory
        .create(erasedRowType)
        .convertToRexNode(watermarkSpec.get.getWatermarkExpr)
    relBuilder.watermark(rowtimeIndex, watermarkRexNode)
  }

  // 4. returns the final RelNode
  relBuilder.build()
}

对Source table进行推断

见CatalogSourceTable.scala

  lazy val tableSource: TableSource[T] = findAndCreateTableSource().asInstanceOf[TableSource[T]]

TableFactoryUtil会根据DDL中的with参数进行推断

private static <T> TableSource<T> findAndCreateTableSource(Map<String, String> properties) {
   try {
      return TableFactoryService
         .find(TableSourceFactory.class, properties)
         .createTableSource(properties);
   } catch (Throwable t) {
      throw new TableException("findAndCreateTableSource failed.", t);
   }
}

主要逻辑在TableFactoryService#filter方法中

private static <T extends TableFactory> List<T> filter(
      List<TableFactory> foundFactories,
      Class<T> factoryClass,
      Map<String, String> properties) {

   Preconditions.checkNotNull(factoryClass);
   Preconditions.checkNotNull(properties);
     //先根据TableSourceFactory.class 过滤出source的factory
   List<T> classFactories = filterByFactoryClass(
      factoryClass,
      properties,
      foundFactories);
     //再根据with参数中的connect.type过滤出满足条件的SourceFactory
   List<T> contextFactories = filterByContext(
      factoryClass,
      properties,
      classFactories);
    //最终判断一下所有properties是否都支持
   return filterBySupportedProperties(
      factoryClass,
      properties,
      classFactories,
      contextFactories);
}

大致调用栈如下

在这里插入图片描述

后续的优化部分其实都是直接基于 RelNode来完成的。

SQL 转换及优化

转换的流程主要分为四个部分，即 1）将 Operation 转换为 RelNode，2）优化 RelNode，3）转换成 ExecNode，4）转换为底层的 Transformation 算子。

如果是DML操作进过SQL转化后会变为ModifyOperation，就会调用Planner的translate方法。

具体实现在PlannerBase.scala

override def translate(
    modifyOperations: util.List[ModifyOperation]): util.List[Transformation[_]] = {
  if (modifyOperations.isEmpty) {
    return List.empty[Transformation[_]]
  }
  // prepare the execEnv before translating
  getExecEnv.configure(
    getTableConfig.getConfiguration,
    Thread.currentThread().getContextClassLoader)
  overrideEnvParallelism()
    //转化为relNode
  val relNodes = modifyOperations.map(translateToRel)
  //SQL优化
  val optimizedRelNodes = optimize(relNodes)
  //转化为execNode
  val execNodes = translateToExecNodePlan(optimizedRelNodes)
  //转化为底层的transfomation
  translateToPlan(execNodes)
}

优化比较复杂，Blink主要是基于Calcite自己的优化，并自定义了一些优化逻辑，有兴趣的读者可以自行研究。

SQL执行

在得到Transformation后的转化逻辑就和streaming模式一致了，可以参考我之前的博客Flink源码阅读之基于Flink1.10的任务提交流程和[Flink源码阅读之基于Flink1.10的任务执行流程](