我使用 Flink 来丰富输入流
case class Input( key: String, message: String )
预先计算的分数
case class Score( key: String, score: Int )
并产生输出
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 )
}
}
这很好用。但是,如果我采取安全点并取消 Flink 作业,则当我从保存点恢复作业时,存储在 ValueState 中的分数会丢失。
据我了解,ScoreEnrichmentFunction 似乎需要使用 CheckPointedFunction 进行扩展
class ScoreEnrichmentFunction extends RichCoFlatMapFunction[Input, Score, Output] with CheckpointedFunction
但我很难理解如何实现方法 snapshotState 和 initializeState 以使用键控状态
override def snapshotState( context: FunctionSnapshotContext ): Unit = ???
override def initializeState( context: FunctionInitializationContext ): Unit = ???
请注意,我使用以下环境:
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 )