当前位置：首页 > news >正文

Apache SeaTunnel Spark引擎执行流程源码分析

news 2025/6/24 6:16:13

1. 任务启动入口

2. 任务执行命令类：SparkTaskExecuteCommand

3. SparkExecution的创建与初始化

3.1 核心组件初始化

3.2 关键对象说明

4. 任务执行：SparkExecution.execute()

5. Source处理流程

5.1 插件初始化

5.2 数据流生成

6. Transform处理流程

6.1 插件初始化

6.2 转换执行

7. Sink处理流程

7.1 插件初始化

7.2 数据输出

执行流程全景图

关键设计总结

本文基于SeaTunnel 2.3.x源码分析Spark引擎执行流程，以seatunnel-examples/seatunnel-spark-connector-v2-example/src/main/java/org/apache/seatunnel/example/spark/v2/SeaTunnelApiExample.java为入口，完整解析Spark引擎的执行流程。

1. 任务启动入口

启动类核心代码：

public static void main(String[] args) {// 1. 创建Spark命令参数对象SparkCommandArgs sparkCommandArgs = new SparkCommandArgs();// 2. 执行SeaTunnel.run()回调Spark执行命令SeaTunnel.run(sparkCommandArgs.buildCommand());
}

buildCommand()返回SparkTaskExecuteCommand实例
SeaTunnel.run()最终调用SparkTaskExecuteCommand.execute()

2. 任务执行命令类：SparkTaskExecuteCommand

核心执行流程：

public void execute() {// 1. 解析配置文件生成Config对象Config config = ConfigBuilder.of(configFile);// 2. 创建SparkExecution实例SparkExecution seaTunnelTaskExecution = new SparkExecution(config);// 3. 执行任务seaTunnelTaskExecution.execute();
}

3. SparkExecution的创建与初始化

3.1 核心组件初始化

public SparkExecution(Config config) {// 创建Spark运行时环境this.sparkRuntimeEnvironment = SparkRuntimeEnvironment.getInstance(config);JobContext jobContext = new JobContext();jobContext.setJobMode(RuntimeEnvironment.getJobMode(config));// 创建三大处理器this.sourcePluginExecuteProcessor = new SourceExecuteProcessor(sparkRuntimeEnvironment, jobContext, config.getConfigList(Constants.SOURCE));this.transformPluginExecuteProcessor = new TransformExecuteProcessor(sparkRuntimeEnvironment, jobContext,TypesafeConfigUtils.getConfigList(config, Constants.TRANSFORM, Collections.emptyList()));this.sinkPluginExecuteProcessor = new SinkExecuteProcessor(sparkRuntimeEnvironment, jobContext, config.getConfigList(Constants.SINK));
}

3.2 关键对象说明

组件	类型	功能
`sourcePluginExecuteProcessor`	`SourceExecuteProcessor`	处理数据源接入
`transformPluginExecuteProcessor`	`TransformExecuteProcessor`	处理数据转换逻辑
`sinkPluginExecuteProcessor`	`SinkExecuteProcessor`	处理数据输出
`sparkRuntimeEnvironment`	`SparkRuntimeEnvironment`	封装SparkSession及运行时环境

4. 任务执行：SparkExecution.execute()

DAG构建流程：

public void execute() throws TaskExecuteException {// 初始化数据集集合List<Dataset<Row>> datasets = new ArrayList<>();// 按顺序执行三大组件datasets = sourcePluginExecuteProcessor.execute(datasets);datasets = transformPluginExecuteProcessor.execute(datasets);sinkPluginExecuteProcessor.execute(datasets);log.info("Spark Execution started");
}

5. Source处理流程

5.1 插件初始化

调用链：

SourceExecuteProcessor()→ super(sparkRuntimeEnvironment, jobContext, sourceConfigs) // 调用父类构造器→ this.plugins = initializePlugins(pluginConfigs)

插件加载核心逻辑：

protected List<SeaTunnelSource<?, ?, ?>> initializePlugins(List<? extends Config> pluginConfigs) {SeaTunnelSourcePluginDiscovery discovery = new SeaTunnelSourcePluginDiscovery();List<SeaTunnelSource<?, ?, ?>> sources = new ArrayList<>();Set<URL> jars = new HashSet<>();for (Config sourceConfig : pluginConfigs) {// 1. 识别插件类型PluginIdentifier identifier = PluginIdentifier.of(ENGINE_TYPE, PLUGIN_TYPE, sourceConfig.getString(PLUGIN_NAME));// 2. 加载依赖JARjars.addAll(discovery.getPluginJarPaths(Lists.newArrayList(identifier)));// 3. 创建插件实例SeaTunnelSource<?, ?, ?> source = discovery.createPluginInstance(identifier);// 4. 初始化插件source.prepare(sourceConfig);source.setJobContext(jobContext);sources.add(source);}// 5. 注册插件JAR到Spark环境sparkRuntimeEnvironment.registerPlugin(new ArrayList<>(jars));return sources;
}

5.2 数据流生成

执行入口：

public List<Dataset<Row>> execute(List<Dataset<Row>> upstreamDataStreams) {List<Dataset<Row>> sources = new ArrayList<>();for (int i = 0; i < plugins.size(); i++) {SeaTunnelSource<?, ?, ?> source = plugins.get(i);Config pluginConfig = pluginConfigs.get(i);// 1. 确定并行度int parallelism = pluginConfig.hasPath(CommonOptions.PARALLELISM.key())? pluginConfig.getInt(CommonOptions.PARALLELISM.key()): sparkRuntimeEnvironment.getSparkConf().getInt(CommonOptions.PARALLELISM.key(), CommonOptions.PARALLELISM.defaultValue());// 2. 创建Dataset（核心步骤）Dataset<Row> dataset = sparkRuntimeEnvironment.getSparkSession().read().format(SeaTunnelSource.class.getSimpleName()) // 使用SeaTunnelSource标识.option(CommonOptions.PARALLELISM.key(), parallelism) // 设置并行度.option(Constants.SOURCE_SERIALIZATION, SerializationUtils.objectToString(source)) // 序列化插件实例.schema((StructType) TypeConverterUtils.convert(source.getProducedType())) // 设置Schema.load();// 3. 注册临时视图registerInputTempView(pluginConfig, dataset);sources.add(dataset);}return sources;
}

临时视图注册逻辑：

void registerInputTempView(Config config, Dataset<Row> dataset) {if (config.hasPath(RESULT_TABLE_NAME)) {String tableName = config.getString(RESULT_TABLE_NAME);// 创建Spark临时视图dataset.createOrReplaceTempView(tableName);}
}

6. Transform处理流程

6.1 插件初始化

关键校验逻辑（在具体Transform实现中）：

public void prepare(Config pluginConfig) {// 必须包含source_table_name和result_table_nameif (!pluginConfig.hasPath(SOURCE_TABLE_NAME) || !pluginConfig.hasPath(RESULT_TABLE_NAME)) {throw new IllegalArgumentException("Missing required table name config");}// 输入输出表名不能相同if (Objects.equals(pluginConfig.getString(SOURCE_TABLE_NAME),pluginConfig.getString(RESULT_TABLE_NAME))) {throw new IllegalArgumentException("Source and result table names must be different");}// 调用具体Transform的配置初始化setConfig(pluginConfig);
}

6.2 转换执行

核心处理流程：

public List<Dataset<Row>> execute(List<Dataset<Row>> upstreamDataStreams) {Dataset<Row> input = upstreamDataStreams.get(0); // 默认使用第一个上游流List<Dataset<Row>> result = new ArrayList<>();for (int i = 0; i < plugins.size(); i++) {SeaTunnelTransform<SeaTunnelRow> transform = plugins.get(i);Config pluginConfig = pluginConfigs.get(i);// 1. 获取输入流（通过source_table_name查找）Dataset<Row> stream = fromSourceTable(pluginConfig, sparkRuntimeEnvironment).orElse(input);// 2. 执行转换input = sparkTransform(transform, stream);// 3. 注册结果表registerInputTempView(pluginConfig, input);result.add(input);}return result;
}

转换算子实现：

private Dataset<Row> sparkTransform(SeaTunnelTransform transform, Dataset<Row> stream) {// 1. 类型系统转换SeaTunnelDataType<?> inputType = TypeConverterUtils.convert(stream.schema());transform.setTypeInfo(inputType);// 2. 创建输出SchemaStructType outputSchema = (StructType) TypeConverterUtils.convert(transform.getProducedType());// 3. 创建行转换器SeaTunnelRowConverter inputConverter = new SeaTunnelRowConverter(inputType);SeaTunnelRowConverter outputConverter = new SeaTunnelRowConverter(transform.getProducedType());// 4. 通过mapPartitions实现转换ExpressionEncoder<Row> encoder = RowEncoder.apply(outputSchema);return stream.mapPartitions((MapPartitionsFunction<Row, Row>) inputIterator -> new TransformIterator( // 自定义迭代器封装转换逻辑inputIterator, transform, outputSchema, inputConverter, outputConverter),encoder).filter(row -> row != null); // 过滤空值
}

自定义迭代器TransformIterator关键逻辑：

public class TransformIterator implements Iterator<Row> {@Overridepublic Row next() {Row row = input.next();// 1. 输入行转SeaTunnelRowSeaTunnelRow inRow = inputConverter.convert(row);// 2. 执行Transform核心逻辑SeaTunnelRow outRow = (SeaTunnelRow) transform.map(inRow);if (outRow != null) {// 3. 输出行转Spark Rowreturn outputConverter.convert(outRow);}return null;}
}

7. Sink处理流程

7.1 插件初始化

特殊处理逻辑：

protected List<SeaTunnelSink<?, ?, ?, ?>> initializePlugins(...) {for (Config sinkConfig : pluginConfigs) {// ... 创建sink实例// 数据保存模式处理if (sink instanceof SupportDataSaveMode) {SupportDataSaveMode saveModeSink = (SupportDataSaveMode) sink;saveModeSink.checkOptions(sinkConfig); // 校验配置}}
}

7.2 数据输出

执行流程：

public List<Dataset<Row>> execute(List<Dataset<Row>> upstreamDataStreams) {Dataset<Row> input = upstreamDataStreams.get(0);for (int i = 0; i < plugins.size(); i++) {Config sinkConfig = pluginConfigs.get(i);SeaTunnelSink<?, ?, ?, ?> sink = plugins.get(i);// 1. 获取输入流Dataset<Row> dataset = fromSourceTable(sinkConfig, sparkRuntimeEnvironment).orElse(input);// 2. 设置类型信息sink.setTypeInfo((SeaTunnelRowType) TypeConverterUtils.convert(dataset.schema()));// 3. 处理数据保存模式if (sink instanceof SupportDataSaveMode) {SupportDataSaveMode saveModeSink = (SupportDataSaveMode) sink;DataSaveMode saveMode = saveModeSink.getDataSaveMode();saveModeSink.handleSaveMode(saveMode);}// 4. 确定并行度int parallelism = sinkConfig.hasPath(CommonOptions.PARALLELISM.key())? sinkConfig.getInt(CommonOptions.PARALLELISM.key()): sparkRuntimeEnvironment.getSparkConf().getInt(CommonOptions.PARALLELISM.key(), CommonOptions.PARALLELISM.defaultValue());// 5. 设置并行度（TODO：当前实现需优化）dataset.sparkSession().read().option(CommonOptions.PARALLELISM.key(), parallelism);// 6. 注入Sink逻辑SparkSinkInjector.inject(dataset.write(), sink).option("checkpointLocation", "/tmp") // TODO：需改为可配置.mode(SaveMode.Append).save();}return null; // Sink是终点，无下游数据流
}

Sink输入流获取逻辑：

protected Optional<Dataset<Row>> fromSourceTable(Config pluginConfig, SparkRuntimeEnvironment env) {if (!pluginConfig.hasPath(SOURCE_TABLE_NAME)) {return Optional.empty();}String sourceTableName = pluginConfig.getString(SOURCE_TABLE_NAME);// 从已注册的临时视图中获取数据集return Optional.of(env.getSparkSession().read().table(sourceTableName));
}

Sink注入器伪代码：

class SparkSinkInjector {static DataFrameWriter inject(DataFrameWriter writer, SeaTunnelSink sink) {// 通过反射或适配器模式将SeaTunnelSink逻辑注入到Spark的DataFrameWriterreturn writer.format(SeaTunnelSink.class.getSimpleName()).option(Constants.SINK_SERIALIZATION, SerializationUtils.objectToString(sink));}
}

执行流程全景图

关键设计总结

插件化架构：
- 通过SPI机制动态加载Source/Transform/Sink插件
- 插件发现机制：SeaTunnel*PluginDiscovery
- 依赖隔离：registerPlugin()管理插件JAR
表驱动DAG：
- source_table_name和result_table_name构成执行链路
- 通过createOrReplaceTempView()实现表注册
- 使用fromSourceTable()实现表查找
类型系统转换：
- TypeConverterUtils处理SeaTunnel类型与Spark类型的双向转换
- SeaTunnelRowConverter实现行级别的数据转换
- 自动Schema推导：source.getProducedType()
执行环境封装：
- SparkRuntimeEnvironment统一管理SparkSession
- 支持批处理和流处理模式
- 集中管理插件依赖和配置
分布式转换：
- Transform阶段使用mapPartitions实现并行处理
- 自定义TransformIterator封装转换逻辑
- 基于Spark的ExpressionEncoder处理类型安全
Sink适配机制：
- 通过SparkSinkInjector桥接SeaTunnelSink和Spark写入体系
- 支持不同的保存模式（Append/Overwrite等）
- 自动处理并行度和检查点配置