本文参考康琪大佬的博客:使Flink SQL Kafka Source支持独立设置并行度

一直觉得 Flink Sql 需要指定算子并行度的功能,哪怕是基于 SQL 解析出来的算子不能添加并行度,source、sink、join 的算子也应该有修改并行度的功能。

恰好看到大佬的博客,Kafka 是最常用的数据源组件了,所以决定在 sqlSubmit 中也加入相应的实现。

Streaming Api 设置并行度

基于 Flink Streaming api,要给 Kafka Source 指定并行度,只需要在 env.addSource() 后面调用 setParallelism() 方法指定并行度就可以,如下:


val kafkaSource = new FlinkKafkaConsumer[ObjectNode](topic, new JsonNodeDeserializationSchema(), Common.getProp)
val stream = env.addSource(kafkaSource)
        .setParallelism(12)

Sql Api 设置并行度

先看一个读kafka 的 SQL

-- kafka source
CREATE TABLE user_log (
  user_id STRING
  ,item_id STRING
  ,category_id STRING
  ,behavior STRING
  ,ts TIMESTAMP(3)
  ,process_time as proctime()
  , WATERMARK FOR ts AS ts - INTERVAL '5' SECOND
) WITH (
  'connector' = 'kafka'
  ,'topic' = 'user_log'
  ,'properties.bootstrap.servers' = 'localhost:9092'
  ,'properties.group.id' = 'user_log'
  ,'scan.startup.mode' = 'latest-offset'
  ,'format' = 'json'
);

CREATE TABLE user_log_sink (
   `day` string
   ,num bigint
   ,min_user_id bigint
   ,max_user_id bigint
) WITH (
   'connector' = 'print'
);

insert into user_log_sink
select `day`
, num
, min_user_id, max_user_id
from(
select DATE_FORMAT(ts,'yyyyMMdd') `day`
,count(distinct user_id) num
,min(cast(replace(user_id,'xxxxxxxxxxxxx','') as bigint)) min_user_id
,max(cast(replace(user_id,'xxxxxxxxxxxxx','') as bigint)) max_user_id
from user_log
-- where DATE_FORMAT(ts,'yyyyMMdd') = date_format(current_timestamp, 'yyyyMMdd')
group by DATE_FORMAT(ts,'yyyyMMdd')
)t1
 where num % 2 = 0
;

流图如下:

仔细看任务流图,所有的算子的并行度都是参数: table.exec.resource.default-parallelism 指定的

要修改 Source 的并行度,其他算子的并行度保持不变,从 Streaming Api 看,只需要给 sql 翻译后的 StreamSource 指定并行度,就可以做到我们想要的结果。

那就直接找到 flink sql 源码 kafka source 创建的地方: KafkaDynamicSource.getScanRuntimeProvider 方法


@Override
public ScanRuntimeProvider getScanRuntimeProvider(ScanContext context) {
    final DeserializationSchema<RowData> keyDeserialization =
            createDeserialization(context, keyDecodingFormat, keyProjection, keyPrefix);

    final DeserializationSchema<RowData> valueDeserialization =
            createDeserialization(context, valueDecodingFormat, valueProjection, null);

    final TypeInformation<RowData> producedTypeInfo =
            context.createTypeInformation(producedDataType);

    final KafkaSource<RowData> kafkaSource =
            createKafkaSource(keyDeserialization, valueDeserialization, producedTypeInfo);

    return new DataStreamScanProvider() {
        @Override
        public DataStream<RowData> produceDataStream(StreamExecutionEnvironment execEnv) {
            if (watermarkStrategy == null) {
                watermarkStrategy = WatermarkStrategy.noWatermarks();
            }
            // 创建 DataStreamSource
            return execEnv.fromSource(
                    kafkaSource, watermarkStrategy, "KafkaSource-" + tableIdentifier);
        }

        @Override
        public boolean isBounded() {
            return kafkaSource.getBoundedness() == Boundedness.BOUNDED;
        }
    };
}

从源码可以看到创建了 KafkaSource ,并且 调用了 execEnv.fromSource 方法,按照 Streaming api 的思路,直接在 execEnv.fromSource 后面添加 setParallelism 就好了,改好的代码如下:


@Override
public ScanRuntimeProvider getScanRuntimeProvider(ScanContext context) {
    final DeserializationSchema<RowData> keyDeserialization =
            createDeserialization(context, keyDecodingFormat, keyProjection, keyPrefix);

    final DeserializationSchema<RowData> valueDeserialization =
            createDeserialization(context, valueDecodingFormat, valueProjection, null);

    final TypeInformation<RowData> producedTypeInfo =
            context.createTypeInformation(producedDataType);

    final KafkaSource<RowData> kafkaSource =
            createKafkaSource(keyDeserialization, valueDeserialization, producedTypeInfo);

    return new DataStreamScanProvider() {
        @Override
        public DataStream<RowData> produceDataStream(StreamExecutionEnvironment execEnv) {
            if (watermarkStrategy == null) {
                watermarkStrategy = WatermarkStrategy.noWatermarks();
            }

            DataStreamSource<RowData> dataDataStreamSource = execEnv.fromSource(
                    kafkaSource, watermarkStrategy, "KafkaSource-" + tableIdentifier);
            int defaultParallelism = execEnv.getParallelism();

            // add by venn for custom source parallelism
            // 很多任务不需要设置并行度,所以加了个判空条件
            // 如果设置的并行度等于 全局的并行度也不做处理
            if (parallelism != null && parallelism > 0 && parallelism != defaultParallelism) {
                dataDataStreamSource.setParallelism(parallelism);
                // todo 参考博客的大佬有断开 source 算子和后续算子的 算子链,我觉得不需要,如果并行度不一样自动会断开
                // 并行度一样,也不需要断开算子链
//                    dataDataStreamSource.disableChaining();
            }
            return dataDataStreamSource;
        }

        @Override
        public boolean isBounded() {
            return kafkaSource.getBoundedness() == Boundedness.BOUNDED;
        }
    };

}

并行度设置需要在 Kafka Table Source 配置项中添加并行度的配置,参考 Flink 自定义 Http Table Source 自定义 source 中获取配置 sql ddl 中的参数的方法:

  1. 在 KafkaConnectorOptions 类中添加配置项
public static final ConfigOption<Integer> SOURCE_PARALLELISM 
    = ConfigOptions.key("source.parallelism")
        .intType()
        .noDefaultValue()
        .withDescription("Defines a custom parallelism for the source. By default, if this option is not defined, the planner will derive the parallelism for each statement individually by also considering the global configuration.");


  1. 在 KafkaDynamicTableFactory 中添加可选参数: SOURCE_PARALLELISM 和 获取参数值并传递到创建 Kafka Source 的地方

添加可选参数不添加,解析 sql DDL 时,无法解析参数


@Override
public Set<ConfigOption<?>> optionalOptions() {
    final Set<ConfigOption<?>> options = new HashSet<>();
    options.add(FactoryUtil.FORMAT);
    options.add(KEY_FORMAT);
    options.add(KEY_FIELDS);
    options.add(KEY_FIELDS_PREFIX);
    options.add(VALUE_FORMAT);
    options.add(VALUE_FIELDS_INCLUDE);
    options.add(TOPIC);
    options.add(TOPIC_PATTERN);
    options.add(PROPS_GROUP_ID);
    options.add(SCAN_STARTUP_MODE);
    options.add(SCAN_STARTUP_SPECIFIC_OFFSETS);
    options.add(SCAN_TOPIC_PARTITION_DISCOVERY);
    options.add(SCAN_STARTUP_TIMESTAMP_MILLIS);
    options.add(SINK_PARTITIONER);
    options.add(SINK_PARALLELISM);
    options.add(DELIVERY_GUARANTEE);
    options.add(TRANSACTIONAL_ID_PREFIX);
    options.add(SINK_SEMANTIC);
    options.add(SOURCE_PARALLELISM);
    return options;
}

解析参数:


@Override
public DynamicTableSource createDynamicTableSource(Context context) {
   
    .....
    // 字段值可能为空,所以类型为 Integer
    final Integer parallelism = tableOptions.getOptional(SOURCE_PARALLELISM).orElse(null);

    return createKafkaTableSource(
            physicalDataType,
            keyDecodingFormat.orElse(null),
            valueDecodingFormat,
            keyProjection,
            valueProjection,
            keyPrefix,
            getSourceTopics(tableOptions),
            getSourceTopicPattern(tableOptions),
            properties,
            startupOptions.startupMode,
            startupOptions.specificOffsets,
            startupOptions.startupTimestampMillis,
            context.getObjectIdentifier().asSummaryString(),
            parallelism);
}

新加一个 createKafkaTableSource 的重载方法,添加 parallelism 变量为参数


protected KafkaDynamicSource createKafkaTableSource(
            DataType physicalDataType,
            @Nullable DecodingFormat<DeserializationSchema<RowData>> keyDecodingFormat,
            DecodingFormat<DeserializationSchema<RowData>> valueDecodingFormat,
            int[] keyProjection,
            int[] valueProjection,
            @Nullable String keyPrefix,
            @Nullable List<String> topics,
            @Nullable Pattern topicPattern,
            Properties properties,
            StartupMode startupMode,
            Map<KafkaTopicPartition, Long> specificStartupOffsets,
            long startupTimestampMillis,
            String tableIdentifier,
            Integer parallelism) {
    return new KafkaDynamicSource(
            physicalDataType,
            keyDecodingFormat,
            valueDecodingFormat,
            keyProjection,
            valueProjection,
            keyPrefix,
            topics,
            topicPattern,
            properties,
            startupMode,
            specificStartupOffsets,
            startupTimestampMillis,
            false,
            tableIdentifier,
            parallelism);
    }

KafkaDynamicSource 构造方法中添加 Integer 类型的变量 parallelism 接收 KafkaDynamicTableFactory 传递过来的 parallelism 参数。

大功告成了吗,测试一下先

全部SQL 同上,在 kafka source 的属性中添加参数: 'source.parallelism' = '2'

  • 注: 全局并行度为 1

CREATE TABLE user_log (
  user_id STRING
  ,item_id STRING
  ,category_id STRING
  ,behavior STRING
  ,ts TIMESTAMP(3)
  ,process_time as proctime()
  , WATERMARK FOR ts AS ts - INTERVAL '5' SECOND
) WITH (
  'connector' = 'kafka'
  ,'topic' = 'user_log'
  ,'properties.bootstrap.servers' = 'localhost:9092'
  ,'properties.group.id' = 'user_log'
  ,'scan.startup.mode' = 'latest-offset'
  ,'format' = 'json'
  ,'source.parallelism' = '2'
);


任务流图:

算子的并行度跟预期的有点不一样,只设置了 source 的并行度,后续的 Calc 的并行度也跟着变了

查看 Calc 算子的源码可以看到:


@Override
protected Transformation<RowData> translateToPlanInternal(PlannerBase planner) {
    final ExecEdge inputEdge = getInputEdges().get(0);
    final Transformation<RowData> inputTransform =
            (Transformation<RowData>) inputEdge.translateToPlan(planner);
    final CodeGeneratorContext ctx =
            new CodeGeneratorContext(planner.getTableConfig())
                    .setOperatorBaseClass(operatorBaseClass);

    final CodeGenOperatorFactory<RowData> substituteStreamOperator =
            CalcCodeGenerator.generateCalcOperator(
                    ctx,
                    inputTransform,
                    (RowType) getOutputType(),
                    JavaScalaConversionUtil.toScala(projection),
                    JavaScalaConversionUtil.toScala(Optional.ofNullable(this.condition)),
                    retainHeader,
                    getClass().getSimpleName());
    return new OneInputTransformation<>(
            inputTransform,
            getDescription(),
            substituteStreamOperator,
            InternalTypeInfo.of(getOutputType()),
            inputTransform.getParallelism());
}

substituteStreamOperator 就是 CodeGen 生产的 Calc 算子,并行度直接取的输入算子的并行度 inputTransform.getParallelism()

所以 Calc 算子的并行度和 Kafka source 的并行度一样,修改源码将并行度设置为 全局并行度


protected Transformation<RowData> translateToPlanInternal(PlannerBase planner) {
    final ExecEdge inputEdge = getInputEdges().get(0);
    final Transformation<RowData> inputTransform =
            (Transformation<RowData>) inputEdge.translateToPlan(planner);
    final CodeGeneratorContext ctx =
            new CodeGeneratorContext(planner.getTableConfig())
                    .setOperatorBaseClass(operatorBaseClass);

    final CodeGenOperatorFactory<RowData> substituteStreamOperator =
            CalcCodeGenerator.generateCalcOperator(
                    ctx,
                    inputTransform,
                    (RowType) getOutputType(),
                    JavaScalaConversionUtil.toScala(projection),
                    JavaScalaConversionUtil.toScala(Optional.ofNullable(this.condition)),
                    retainHeader,
                    getClass().getSimpleName());

    // add by venn for custom source parallelism
    int parallelism = inputTransform.getParallelism();
    int defaultParallelism = ExecutionConfig.PARALLELISM_DEFAULT;
    if (parallelism != defaultParallelism) {
        // update calc operator parallelism to default parallelism
        parallelism = defaultParallelism;
    }

    return new OneInputTransformation<>(
            inputTransform,
            getDescription(),
            substituteStreamOperator,
            InternalTypeInfo.of(getOutputType()),
            parallelism);
}

生成完的算子流图:

现在才大功告成

完整代码参考: github sqlSubmit

-----------------------分割线------------------------------------

  • 2022-01-13 改

在 sqlSubmit 的配置文件中添加了参数: table.exec.source.force-break-chain ,用以在 kafka sql source 后面判断是否要断开下游算子和 source 的 chain。

由于在 CommonExecCalc 拿不到 ddl 中定义的参数,添加了一个任务级别的参数( 在 sqlSubmit.properties 中添加参数 table.exec.source.force-break-chain 默认值为 false,如果为 true 会断开 chain),通过 env 放到 GlobalJobParameter 中,KafkaDynamicSource 通过 env 从 GlobalJobParameter 中取,CommonExecCalc 有 planner 可以直接拿到任务级别的参数。

  • 对于指定了 kafka sql source 的情况,是否需要断开 source 算子和后续的算子:根据 source 的并行度和默认并行度判断是否断开,还是用参数控制是否断开。最后还是决定安装大佬博客里的写的,用参数控制是否断开。主要是source 和后续算子的关系不定,如果只是基于并行度判断是否 chain 在一起太过僵硬,还不如再加一个参数控制,这样默认算子会 chain 在一起,下游算子和 source 并行度保持一致,如果需要端口,后续算子的并行度会和全局并行度保持一致。

欢迎关注Flink菜鸟公众号,会不定期更新Flink(开发技术)相关的推文

posted on 2022-01-08 15:44  Flink菜鸟  阅读(5314)  评论(0编辑  收藏  举报