我使用flink来丰富输入流

case class Input( key: String, message: String )

用预先计算的分数

case class Score( key: String, score: Int )

并产生一个输出

case class Output( key: String, message: String, score: Int )

输入流和分数流都从Kafka主题中读取，结果输出流也发布到Kafka

val processed = inputStream.connect( scoreStream )
                           .flatMap( new ScoreEnrichmentFunction )
                           .addSink( producer )

使用以下scoreenrichmentfunction：

class ScoreEnrichmentFunction extends RichCoFlatMapFunction[Input, Score, Output]
{
    val scoreStateDescriptor = new ValueStateDescriptor[Score]( "saved scores", classOf[Score] )
    lazy val scoreState: ValueState[Score] = getRuntimeContext.getState( scoreStateDescriptor )

    override def flatMap1( input: Input, out: Collector[Output] ): Unit = 
    {
        Option( scoreState.value ) match {
            case None => out.collect( Output( input.key, input.message, -1 ) )
            case Some( score ) => out.collect( Output( input.key, input.message, score.score ) )  
        }
    }

    override def flatMap2( score: Score, out: Collector[Output] ): Unit = 
    {
        scoreState.update( score )
    } 
}

这很有效。但是，如果我选择一个安全点并取消flink作业，当我从保存点恢复作业时，存储在valuestate中的分数将丢失。
据我所知，scoreenrichmentfunction似乎需要使用checkpointedfunction进行扩展

class ScoreEnrichmentFunction extends RichCoFlatMapFunction[Input, Score, Output] with CheckpointedFunction

但是我很难理解如何实现snapshotstate和initializestate方法来处理键控状态

override def snapshotState( context: FunctionSnapshotContext ): Unit = ???

override def initializeState( context: FunctionInitializationContext ): Unit = ???

请注意，我使用以下env：

val env = StreamExecutionEnvironment.getExecutionEnvironment
    env.setParallelism( 2 )
    env.setBufferTimeout( 1 )
    env.enableCheckpointing( 1000 )
    env.getCheckpointConfig.enableExternalizedCheckpoints( ExternalizedCheckpointCleanup.RETAIN_ON_CANCELLATION )
    env.getCheckpointConfig.setCheckpointingMode( CheckpointingMode.EXACTLY_ONCE )
    env.getCheckpointConfig.setMinPauseBetweenCheckpoints( 500 )
    env.getCheckpointConfig.setCheckpointTimeout( 60000 )
    env.getCheckpointConfig.setFailOnCheckpointingErrors( false )
    env.getCheckpointConfig.setMaxConcurrentCheckpoints( 1 )