Spark Streaming + Kafka direct 从Zookeeper中恢复offset

在上一遍《将 Spark Streaming + Kafka direct 的 offset 保存进入Zookeeper》中，我们已经成功的将 topic 的 partition 的 offset 保存到了 Zookeeper中，使监控工具发挥了其监控效果。那现在是时候来处理《“Spark Streaming + Kafka direct + checkpoints + 代码改变” 引发的问题》中提到的问题了。

---

解决的方法是：分别从Kafka中获得某个Topic当前每个partition的offset，再从Zookeeper中获得某个consumer消费当前Topic中每个partition的offset，最后再这两个根据项目情况进行合并，就可以了。

一、具体实现

1、程序实现，如下：

public class SparkStreamingOnKafkaDirect{

    public static JavaStreamingContext createContext(){

        SparkConf conf = new SparkConf().setMaster("local[4]").setAppName("SparkStreamingOnKafkaDirect");

        JavaStreamingContext jsc = new JavaStreamingContext(conf, Durations.seconds(30));
        jsc.checkpoint("/checkpoint");

        Map kafkaParams = new HashMap();
        kafkaParams.put("metadata.broker.list","192.168.1.151:1234,192.168.1.151:1235,192.168.1.151:1236");

        Map topicOffsets = getTopicOffsets("192.168.1.151:1234,192.168.1.151:1235,192.168.1.151:1236", "kafka_direct");

        Map consumerOffsets = getConsumerOffsets("192.168.1.151:2181", "spark-group", "kafka_direct");
        if(null!=consumerOffsets && consumerOffsets.size()>0){
            topicOffsets.putAll(consumerOffsets);
        }

//        for(Map.Entry item:topicOffsets.entrySet()){
//            item.setValue(0l);
//        }

      for(Map.Entry entry:topicOffsets.entrySet()){
          System.out.println(entry.getKey().topic()+"\t"+entry.getKey().partition()+"\t"+entry.getValue());
      }

        JavaInputDStream lines = KafkaUtils.createDirectStream(jsc,
                String.class, String.class, StringDecoder.class,
                StringDecoder.class, String.class, kafkaParams,
                topicOffsets, new Function,String>() {

                    public String call(MessageAndMetadata v1)
                            throws Exception {
                        return v1.message();
                    }
                });



        final AtomicReference offsetRanges = new AtomicReference<>();

        JavaDStream words = lines.transform(
                new Function, JavaRDD>() {
                    @Override
                    public JavaRDD call(JavaRDD rdd) throws Exception {
                      OffsetRange[] offsets = ((HasOffsetRanges) rdd.rdd()).offsetRanges();
                      offsetRanges.set(offsets);
                      return rdd;
                    }
                  }
                ).flatMap(new FlatMapFunction() {
                    public Iterable call(
                           String event)
                            throws Exception {
                        return Arrays.asList(event);
                    }
                });

        JavaPairDStream pairs = words
                .mapToPair(new PairFunction() {

                    public Tuple2 call(
                            String word) throws Exception {
                        return new Tuple2(
                                word, 1);
                    }
                });

        JavaPairDStream wordsCount = pairs
                .reduceByKey(new Function2() {
                    public Integer call(Integer v1, Integer v2)
                            throws Exception {
                        return v1 + v2;
                    }
                });

        lines.foreachRDD(new VoidFunction>(){
            @Override
            public void call(JavaRDD t) throws Exception {

                ObjectMapper objectMapper = new ObjectMapper();

                CuratorFramework  curatorFramework = CuratorFrameworkFactory.builder()
                        .connectString("192.168.1.151:2181").connectionTimeoutMs(1000)
                        .sessionTimeoutMs(10000).retryPolicy(new RetryUntilElapsed(1000, 1000)).build();

                curatorFramework.start();

                for (OffsetRange offsetRange : offsetRanges.get()) {
                    final byte[] offsetBytes = objectMapper.writeValueAsBytes(offsetRange.untilOffset());
                    String nodePath = "/consumers/spark-group/offsets/" + offsetRange.topic()+ "/" + offsetRange.partition();
                    if(curatorFramework.checkExists().forPath(nodePath)!=null){
                            curatorFramework.setData().forPath(nodePath,offsetBytes);
                        }else{
                            curatorFramework.create().creatingParentsIfNeeded().forPath(nodePath, offsetBytes);
                        }
                }

                curatorFramework.close();
            }

        });

        wordsCount.print();

        return jsc;
    }


    public static Map getConsumerOffsets(String zkServers, 
                String groupID, String topic) { 
        Map retVals = new HashMap();

        ObjectMapper objectMapper = new ObjectMapper();
        CuratorFramework  curatorFramework = CuratorFrameworkFactory.builder()
                .connectString(zkServers).connectionTimeoutMs(1000)
                .sessionTimeoutMs(10000).retryPolicy(new RetryUntilElapsed(1000, 1000)).build();

        curatorFramework.start();

        try{
        String nodePath = "/consumers/"+groupID+"/offsets/" + topic;
        if(curatorFramework.checkExists().forPath(nodePath)!=null){
            List partitions=curatorFramework.getChildren().forPath(nodePath);
            for(String partiton:partitions){
                int partitionL=Integer.valueOf(partiton);
                Long offset=objectMapper.readValue(curatorFramework.getData().forPath(nodePath+"/"+partiton),Long.class);
                TopicAndPartition topicAndPartition=new TopicAndPartition(topic,partitionL);
                retVals.put(topicAndPartition, offset);
            }
        }
        }catch(Exception e){
            e.printStackTrace();
        }
        curatorFramework.close();

        return retVals;
    } 

    public static Map getTopicOffsets(String zkServers, String topic){
        Map retVals = new HashMap();

        for(String zkServer:zkServers.split(",")){
        SimpleConsumer simpleConsumer = new SimpleConsumer(zkServer.split(":")[0], 
                Integer.valueOf(zkServer.split(":")[1]), 
                10000, 
                1024, 
                "consumer"); 
        TopicMetadataRequest topicMetadataRequest = new TopicMetadataRequest(Arrays.asList(topic));
        TopicMetadataResponse topicMetadataResponse = simpleConsumer.send(topicMetadataRequest);

        for (TopicMetadata metadata : topicMetadataResponse.topicsMetadata()) {
            for (PartitionMetadata part : metadata.partitionsMetadata()) {
                Broker leader = part.leader();
                if (leader != null) { 
                    TopicAndPartition topicAndPartition = new TopicAndPartition(topic, part.partitionId()); 

                    PartitionOffsetRequestInfo partitionOffsetRequestInfo = new PartitionOffsetRequestInfo(kafka.api.OffsetRequest.LatestTime(), 10000); 
                    OffsetRequest offsetRequest = new OffsetRequest(ImmutableMap.of(topicAndPartition, partitionOffsetRequestInfo), kafka.api.OffsetRequest.CurrentVersion(), simpleConsumer.clientId()); 
                    OffsetResponse offsetResponse = simpleConsumer.getOffsetsBefore(offsetRequest); 

                    if (!offsetResponse.hasError()) { 
                        long[] offsets = offsetResponse.offsets(topic, part.partitionId()); 
                        retVals.put(topicAndPartition, offsets[0]);
                    }
                }
            }
        }
        simpleConsumer.close();
        }
        return retVals;
    }

    public static void main(String[] args)  throws Exception{
        JavaStreamingContextFactory factory = new JavaStreamingContextFactory() {
            public JavaStreamingContext create() {
              return createContext();
            }
          };

        JavaStreamingContext jsc = JavaStreamingContext.getOrCreate("/checkpoint", factory);

        jsc.start();

        jsc.awaitTermination();
        jsc.close();

    }

}

2、准备测试环境，并记录目前consumer中的信息，如下图：

从界面上可以看到，目前所有的消息都已经被处理过了。
现在向kafka_direct中新增一个消息，如下图：

3、运行Spark Streaming 程序(注意：要先清空 checkpoint 目录下的内容)，观察命令行输出情况，及kafka manager中关于 spark-group的变化情况：
命令行输出：

打印出了，从zookeeper中读取到的offset。

打印出了，从Kafka的kafka_direct中消费的消息的结果数据。

从图片中，可以看到consumer offset 和 logSize 是一样的。

4、下面我们人为的将topic的partition的offset的值设置为0，看其是否会打印出所有消息的结果数据。
（取消上面程序中注释的部分即可）

5、再次运行Spark Streaming 程序(注意：要先清空 checkpoint 目录下的内容)，看命令行输出效果：

从图中，可以看出Spark Streaming 程序将之前所有的测试消息都重新处理了一次。

至此，《“Spark Streaming + Kafka direct + checkpoints + 代码改变” 引发的问题》的整个解决的过程都已经结束了。

说明：源代码中，由于时间紧，没有写注释，等后面有时间了再补上。