1

对于大量 Flink SQL 查询(以下 100 条),Flink 命令行客户端在 Yarn 集群上失败并显示“JobManager 在 600000 毫秒内没有响应”,即集群上永远不会启动作业。

  • 在最后一个 TaskManager 启动后,JobManager 日志除了“在 JobManager 中找不到 ID 为 5cd95f89ed7a66ec44f2d19eca0592f7 的作业”的 DEBUG 日志之外什么都没有,表明它可能卡住了(创建 ExecutionGraph?)。
  • 与本地的独立 java 程序相同(最初是高 CPU)
  • 注意: structStream 中的每一行包含 515 列(许多最终为空),其中包括一个包含原始消息的列。
  • 在 YARN 集群中,我们为 TaskManager 指定 18GB,为 JobManager 指定 18GB,每个插槽 5 个,并行度为 725(我们的 Kafka 源中的分区)。

Flink SQL 查询:

select count (*), 'idnumber' as criteria, Environment, CollectedTimestamp, 
       EventTimestamp, RawMsg, Source 
from structStream
where Environment='MyEnvironment' and Rule='MyRule' and LogType='MyLogType' 
      and Outcome='Success'
group by tumble(proctime, INTERVAL '1' SECOND), Environment, 
         CollectedTimestamp, EventTimestamp, RawMsg, Source

代码

public static void main(String[] args) throws Exception {
    FileSystems.newFileSystem(KafkaReadingStreamingJob.class
                             .getResource(WHITELIST_CSV).toURI(), new HashMap<>());

    final StreamExecutionEnvironment streamingEnvironment = getStreamExecutionEnvironment();
    final StreamTableEnvironment tableEnv = TableEnvironment.getTableEnvironment(streamingEnvironment);

    final DataStream<Row> structStream = getKafkaStreamOfRows(streamingEnvironment);
    tableEnv.registerDataStream("structStream", structStream);
    tableEnv.scan("structStream").printSchema();

    for (int i = 0; i < 100; i++) {
        for (String query : Queries.sample) {
            // Queries.sample has one query that is above. 
            Table selectQuery = tableEnv.sqlQuery(query);

            DataStream<Row> selectQueryStream =                                                 
                               tableEnv.toAppendStream(selectQuery, Row.class);
            selectQueryStream.print();
        }
    }

    // execute program
    streamingEnvironment.execute("Kafka Streaming SQL");
}

private static DataStream<Row> getKafkaStreamOfRows(StreamExecutionEnvironment environment) throws Exception {
    Properties properties = getKafkaProperties();

    // TestDeserializer deserializes the JSON to a ROW of string columns (515)
    // and also adds a column for the raw message. 
    FlinkKafkaConsumer011 consumer = new         
         FlinkKafkaConsumer011(KAFKA_TOPIC_TO_CONSUME, new TestDeserializer(getRowTypeInfo()), properties);
    DataStream<Row> stream = environment.addSource(consumer);

    return stream;
}

private static RowTypeInfo getRowTypeInfo() throws Exception {
    // This has 515 fields. 
    List<String> fieldNames = DDIManager.getDDIFieldNames();
    fieldNames.add("rawkafka"); // rawMessage added by TestDeserializer
    fieldNames.add("proctime");

    // Fill typeInformationArray with StringType to all but the last field which is of type Time
    .....
    return new RowTypeInfo(typeInformationArray, fieldNamesArray);
}

private static StreamExecutionEnvironment getStreamExecutionEnvironment() throws IOException {
    final StreamExecutionEnvironment env =                      
    StreamExecutionEnvironment.getExecutionEnvironment(); 
    env.setStreamTimeCharacteristic(TimeCharacteristic.ProcessingTime);

    env.enableCheckpointing(60000);
    env.setStateBackend(new FsStateBackend(CHECKPOINT_DIR));
    env.setParallelism(725);
    return env;
}

private static DataStream<Row> getKafkaStreamOfRows(StreamExecutionEnvironment environment) throws Exception {
    Properties properties = getKafkaProperties();

    // TestDeserializer deserializes the JSON to a ROW of string columns (515)
    // and also adds a column for the raw message. 
    FlinkKafkaConsumer011 consumer = new FlinkKafkaConsumer011(KAFKA_TOPIC_TO_CONSUME, new  TestDeserializer(getRowTypeInfo()), properties);
    DataStream<Row> stream = environment.addSource(consumer);

    return stream;
}

private static RowTypeInfo getRowTypeInfo() throws Exception {
    // This has 515 fields. 
    List<String> fieldNames = DDIManager.getDDIFieldNames();
    fieldNames.add("rawkafka"); // rawMessage added by TestDeserializer
    fieldNames.add("proctime");

    // Fill typeInformationArray with StringType to all but the last field which is of type Time
    .....
    return new RowTypeInfo(typeInformationArray, fieldNamesArray);
}

private static StreamExecutionEnvironment getStreamExecutionEnvironment() throws IOException {
    final StreamExecutionEnvironment env =     StreamExecutionEnvironment.getExecutionEnvironment(); 
    env.setStreamTimeCharacteristic(TimeCharacteristic.ProcessingTime);

    env.enableCheckpointing(60000);
    env.setStateBackend(new FsStateBackend(CHECKPOINT_DIR));
    env.setParallelism(725);
    return env;
}
4

1 回答 1

0

这在我看来好像 JobManager 因太多并发运行的作业而超载。我建议将作业分配给更多的 JobManagers / Flink 集群。

于 2018-04-16T12:17:23.923 回答