一个流应用程序必须全天候运行，所有必须能够解决应用程序逻辑无关的故障（如系统错误，JVM崩溃等）。为了使这成为可能，Spark Streaming需要checkpoint足够的信息到容错存储系统中， 以使系统从故障中恢复。
元数据checkpoint主要是为了从driver故障中恢复数据。如果transformation操作被用到了，数据checkpoint即使在简单的操作中都是必须的。
def createContext(checkpointDirectory: String)
: StreamingContext = {
// If you do not see this printed, that means the StreamingContext has been loaded
// from the new checkpoint
println("Creating new context")
val sparkConf = new SparkConf().setAppName("DynamicRange")
// Create the context with a 1 second batch size
val ssc = new StreamingContext(sparkConf, Seconds(8))
ssc.checkpoint(checkpointDirectory)
...
//你的kafka streaming的相关代码最好放在这里，不然有可能抛异常：spark checkpoint KafkaInputDStream has not been initialize
//create kafka stream
val fullLines = KafkaUtils.createStream(ssc, SystemConfig.config.kafkaZkQuorum, SystemConfig.config.kafkaGroup, topicMap);
//parse data string
val valueLines = fullLines.map(_._2)
..
ssc
}
def main(args: Array[String]) {
var ssc: StreamingContext = null
try {
ssc = StreamingContext.getOrCreate(".", () => {
println("get context fail, try to create a new one.")
createContext(".")
})
} catch{
case e:Exception =>{
println("get context exception, try to create a new one.")
ssc = createContext(".")
}
}
ssc.start()
ssc.awaitTermination()
}
Spark Streaming metadata checkpoint
原文：http://www.cnblogs.com/hd-zg/p/5831235.html